Fix lint issues

operator prices, delivered via cost_meter middleware config

  Operator-pinned model prices (ProviderModel.InputPer1k/OutputPer1k) were
  stored and round-tripped through the API but never consumed for billing:
  the proxy priced exclusively from a defaults_pricing.yaml embedded in
  its binary. Any model outside that file recorded $0 spend, price changes
  required a proxy release, and the catalog and the YAML were two
  hand-synced copies of the same rates, guarded only by a manually
  maintained coverage list that had already drifted.

.goreleaser_ui.yaml

@@ -24,6 +24,8 @@ builds:
     ldflags:
       - -s -w -X github.com/netbirdio/netbird/version.version={{.Version}} -X main.commit={{.Commit}} -X main.date={{.CommitDate}} -X main.builtBy=goreleaser
     mod_timestamp: "{{ .CommitTimestamp }}"
+    tags:
+      - production
 
   - id: netbird-ui-windows-amd64
     dir: client/ui
@@ -39,6 +41,8 @@ builds:
       - -s -w -X github.com/netbirdio/netbird/version.version={{.Version}} -X main.commit={{.Commit}} -X main.date={{.CommitDate}} -X main.builtBy=goreleaser
       - -H windowsgui
     mod_timestamp: "{{ .CommitTimestamp }}"
+    tags:
+      - production
 
   - id: netbird-ui-windows-arm64
     dir: client/ui
@@ -55,6 +59,8 @@ builds:
       - -s -w -X github.com/netbirdio/netbird/version.version={{.Version}} -X main.commit={{.Commit}} -X main.date={{.CommitDate}} -X main.builtBy=goreleaser
       - -H windowsgui
     mod_timestamp: "{{ .CommitTimestamp }}"
+    tags:
+      - production
 
 archives:
   - id: linux-arch

.goreleaser_ui_darwin.yaml

@@ -29,6 +29,8 @@ builds:
     ldflags:
       - -s -w -X github.com/netbirdio/netbird/version.version={{.Version}} -X main.commit={{.Commit}} -X main.date={{.CommitDate}} -X main.builtBy=goreleaser
     mod_timestamp: "{{ .CommitTimestamp }}"
+    tags:
+      - production
 
 universal_binaries:
   - id: netbird-ui-darwin

client/android/client.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"os"
 	"slices"
+	"strings"
 	"sync"
 	"time"
 
@@ -299,6 +300,13 @@ func (c *Client) SetInfoLogLevel() {
 // PeersList return with the list of the PeerInfos
 func (c *Client) PeersList() *PeerInfoArray {
 
+	// The recorder only caches transfer counters and handshake times; nothing
+	// refreshes them on its own, so without this they read as zero. The desktop
+	// daemon does the same before serving a full peer status.
+	if err := c.recorder.RefreshWireGuardStats(); err != nil {
+		log.Debugf("failed to refresh WireGuard stats: %v", err)
+	}
+
 	fullStatus := c.recorder.GetFullStatus()
 
 	peerInfos := make([]PeerInfo, len(fullStatus.Peers))
@@ -309,6 +317,20 @@ func (c *Client) PeersList() *PeerInfoArray {
 			FQDN:       p.FQDN,
 			ConnStatus: int(p.ConnStatus),
 			Routes:     PeerRoutes{routes: maps.Keys(p.GetRoutes())},
+
+			PubKey:                     p.PubKey,
+			Latency:                    formatDuration(p.Latency),
+			LatencyMs:                  p.Latency.Milliseconds(),
+			BytesRx:                    p.BytesRx,
+			BytesTx:                    p.BytesTx,
+			ConnStatusUpdate:           formatTime(p.ConnStatusUpdate),
+			Relayed:                    p.Relayed,
+			RosenpassEnabled:           p.RosenpassEnabled,
+			LastWireguardHandshake:     formatTime(p.LastWireguardHandshake),
+			LocalIceCandidateType:      p.LocalIceCandidateType,
+			RemoteIceCandidateType:     p.RemoteIceCandidateType,
+			LocalIceCandidateEndpoint:  p.LocalIceCandidateEndpoint,
+			RemoteIceCandidateEndpoint: p.RemoteIceCandidateEndpoint,
 		}
 		peerInfos[n] = pi
 	}
@@ -439,10 +461,6 @@ func (c *Client) RemoveConnectionListener() {
 	c.recorder.RemoveConnectionListener()
 }
 
-func (c *Client) toggleRoute(command routeCommand) error {
-	return command.toggleRoute()
-}
-
 func (c *Client) getRouteManager() (routemanager.Manager, error) {
 	client := c.getConnectClient()
 	if client == nil {
@@ -462,22 +480,22 @@ func (c *Client) getRouteManager() (routemanager.Manager, error) {
 	return manager, nil
 }
 
-func (c *Client) SelectRoute(route string) error {
+func (c *Client) SelectRoute(id string) error {
 	manager, err := c.getRouteManager()
 	if err != nil {
 		return err
 	}
 
-	return c.toggleRoute(selectRouteCommand{route: route, manager: manager})
+	return manager.SelectRoutes([]route.NetID{route.NetID(id)}, true)
 }
 
-func (c *Client) DeselectRoute(route string) error {
+func (c *Client) DeselectRoute(id string) error {
 	manager, err := c.getRouteManager()
 	if err != nil {
 		return err
 	}
 
-	return c.toggleRoute(deselectRouteCommand{route: route, manager: manager})
+	return manager.DeselectRoutes([]route.NetID{route.NetID(id)})
 }
 
 // getNetworkDomainsFromRoute extracts domains from a route and enriches each domain
@@ -512,3 +530,28 @@ func exportEnvList(list *EnvList) {
 		}
 	}
 }
+
+// formatDuration renders a duration for display, trimming the fractional part
+// to two digits so latencies read as "12.34ms" rather than "12.345678ms".
+func formatDuration(d time.Duration) string {
+	ds := d.String()
+	dotIndex := strings.Index(ds, ".")
+	if dotIndex == -1 {
+		return ds
+	}
+
+	endIndex := min(dotIndex+3, len(ds))
+
+	// Skip the remaining digits so only the unit suffix is appended back.
+	unitStart := endIndex
+	for unitStart < len(ds) && ds[unitStart] >= '0' && ds[unitStart] <= '9' {
+		unitStart++
+	}
+	return ds[:endIndex] + ds[unitStart:]
+}
+
+// formatTime renders a timestamp in UTC using a fixed layout. The zero time is
+// passed through as-is so the UI can recognise it and show "never" instead.
+func formatTime(t time.Time) string {
+	return t.UTC().Format("2006-01-02 15:04:05")
+}

client/android/peer_notifier.go

@@ -12,12 +12,30 @@ const (
 )
 
 // PeerInfo describe information about the peers. It designed for the UI usage
+//
+// The fields below ConnStatus back the peer detail screen. Durations and times
+// are pre-formatted into strings so the UI does not have to know Go's layouts;
+// Latency is additionally exposed as LatencyMs for colour coding.
 type PeerInfo struct {
 	IP         string
 	IPv6       string
 	FQDN       string
 	ConnStatus int
 	Routes     PeerRoutes
+
+	PubKey                     string
+	Latency                    string
+	LatencyMs                  int64
+	BytesRx                    int64
+	BytesTx                    int64
+	ConnStatusUpdate           string
+	Relayed                    bool
+	RosenpassEnabled           bool
+	LastWireguardHandshake     string
+	LocalIceCandidateType      string
+	RemoteIceCandidateType     string
+	LocalIceCandidateEndpoint  string
+	RemoteIceCandidateEndpoint string
 }
 
 func (p *PeerInfo) GetPeerRoutes() *PeerRoutes {

client/android/profile_manager.go

@@ -189,6 +189,19 @@ func (pm *ProfileManager) LogoutProfile(id string) error {
 	return nil
 }
 
+// RenameProfile changes a profile's display name. The profile ID, and therefore
+// its on-disk filename, is left untouched: only the "name" field of the config
+// is rewritten. This works for the default profile too, whose config lives in
+// netbird.cfg rather than under profiles/.
+func (pm *ProfileManager) RenameProfile(id string, newName string) error {
+	if err := pm.serviceMgr.RenameProfile(profilemanager.ID(id), androidUsername, newName); err != nil {
+		return fmt.Errorf("failed to rename profile: %w", err)
+	}
+
+	log.Infof("renamed profile %s to: %s", id, newName)
+	return nil
+}
+
 // RemoveProfile deletes a profile
 func (pm *ProfileManager) RemoveProfile(id string) error {
 	// Use ServiceManager (removes profile from profiles/ directory)

client/android/route_command.go

@@ -1,70 +0,0 @@
-//go:build android
-
-package android
-
-import (
-	"fmt"
-
-	log "github.com/sirupsen/logrus"
-	"golang.org/x/exp/maps"
-
-	"github.com/netbirdio/netbird/client/internal/routemanager"
-	"github.com/netbirdio/netbird/route"
-)
-
-func executeRouteToggle(id string, manager routemanager.Manager,
-	operationName string,
-	routeOperation func(routes []route.NetID, allRoutes []route.NetID) error) error {
-	netID := route.NetID(id)
-	routes := []route.NetID{netID}
-
-	routesMap := manager.GetClientRoutesWithNetID()
-	routes = route.ExpandV6ExitPairs(routes, routesMap)
-
-	log.Debugf("%s with ids: %v", operationName, routes)
-
-	if err := routeOperation(routes, maps.Keys(routesMap)); err != nil {
-		log.Debugf("error when %s: %s", operationName, err)
-		return fmt.Errorf("error %s: %w", operationName, err)
-	}
-
-	manager.TriggerSelection(manager.GetClientRoutes())
-
-	return nil
-}
-
-type routeCommand interface {
-	toggleRoute() error
-}
-
-type selectRouteCommand struct {
-	route   string
-	manager routemanager.Manager
-}
-
-func (s selectRouteCommand) toggleRoute() error {
-	routeSelector := s.manager.GetRouteSelector()
-	if routeSelector == nil {
-		return fmt.Errorf("no route selector available")
-	}
-
-	routeOperation := func(routes []route.NetID, allRoutes []route.NetID) error {
-		return routeSelector.SelectRoutes(routes, true, allRoutes)
-	}
-
-	return executeRouteToggle(s.route, s.manager, "selecting route", routeOperation)
-}
-
-type deselectRouteCommand struct {
-	route   string
-	manager routemanager.Manager
-}
-
-func (d deselectRouteCommand) toggleRoute() error {
-	routeSelector := d.manager.GetRouteSelector()
-	if routeSelector == nil {
-		return fmt.Errorf("no route selector available")
-	}
-
-	return executeRouteToggle(d.route, d.manager, "deselecting route", routeSelector.DeselectRoutes)
-}

client/internal/conn_mgr.go

@@ -385,11 +385,20 @@ func inactivityThresholdEnv() *time.Duration {
 		return nil
 	}
 
-	parsedMinutes, err := strconv.Atoi(envValue)
-	if err != nil || parsedMinutes <= 0 {
-		return nil
+	// Documented format: a Go duration such as "30m" or "1h".
+	if d, err := time.ParseDuration(envValue); err == nil {
+		if d <= 0 {
+			return nil
+		}
+		return &d
 	}
 
-	d := time.Duration(parsedMinutes) * time.Minute
-	return &d
+	// Backwards compatibility: a bare integer used to be interpreted as minutes.
+	if parsedMinutes, err := strconv.Atoi(envValue); err == nil && parsedMinutes > 0 {
+		d := time.Duration(parsedMinutes) * time.Minute
+		return &d
+	}
+
+	log.Warnf("invalid %s value %q: expected a Go duration such as 30m or 1h", lazyconn.EnvInactivityThreshold, envValue)
+	return nil
 }

client/internal/conn_mgr_test.go

@@ -104,3 +104,38 @@ func TestConnMgr_ActivatePeerConcurrentWithLifecycle(t *testing.T) {
 	close(done)
 	wg.Wait()
 }
+
+func TestInactivityThresholdEnv(t *testing.T) {
+	tests := []struct {
+		name string
+		val  string
+		want *time.Duration
+	}{
+		{name: "unset", val: "", want: nil},
+		{name: "go duration minutes", val: "30m", want: durPtr(30 * time.Minute)},
+		{name: "go duration hours", val: "1h", want: durPtr(time.Hour)},
+		{name: "go duration seconds", val: "90s", want: durPtr(90 * time.Second)},
+		{name: "bare integer is minutes (backwards compat)", val: "5", want: durPtr(5 * time.Minute)},
+		{name: "zero duration", val: "0s", want: nil},
+		{name: "zero integer", val: "0", want: nil},
+		{name: "negative duration", val: "-5m", want: nil},
+		{name: "garbage", val: "abc", want: nil},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv(lazyconn.EnvInactivityThreshold, tc.val)
+			got := inactivityThresholdEnv()
+			switch {
+			case tc.want == nil && got != nil:
+				t.Fatalf("want nil, got %v", *got)
+			case tc.want != nil && got == nil:
+				t.Fatalf("want %v, got nil", *tc.want)
+			case tc.want != nil && *got != *tc.want:
+				t.Fatalf("want %v, got %v", *tc.want, *got)
+			}
+		})
+	}
+}
+
+func durPtr(d time.Duration) *time.Duration { return &d }

client/internal/connect.go

@@ -34,6 +34,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/internal/statemanager"
 	"github.com/netbirdio/netbird/client/internal/stdnet"
+	"github.com/netbirdio/netbird/client/internal/tunnelnotifier"
 	"github.com/netbirdio/netbird/client/internal/updater"
 	"github.com/netbirdio/netbird/client/internal/updater/installer"
 	nbnet "github.com/netbirdio/netbird/client/net"
@@ -136,10 +137,13 @@ func (c *ConnectClient) RunOniOS(
 	// Set GC percent to 5% to reduce memory usage as iOS only allows 50MB of memory for the extension.
 	debug.SetGCPercent(5)
 
+	notifier := tunnelnotifier.New(networkChangeListener, dnsManager)
+	defer notifier.Close()
+
 	mobileDependency := MobileDependency{
 		FileDescriptor:        fileDescriptor,
-		NetworkChangeListener: networkChangeListener,
-		DnsManager:            dnsManager,
+		NetworkChangeListener: notifier,
+		DnsManager:            notifier,
 		StateFilePath:         stateFilePath,
 		TempDir:               cacheDir,
 	}

client/internal/mobile_dependency.go

@@ -11,12 +11,14 @@ import (
 
 // MobileDependency collect all dependencies for mobile platform
 type MobileDependency struct {
-	// Android only
-	TunAdapter            device.TunAdapter
-	IFaceDiscover         stdnet.ExternalIFaceDiscover
+	// Android and iOS
 	NetworkChangeListener listener.NetworkChangeListener
-	HostDNSAddresses      []netip.AddrPort
-	DnsReadyListener      dns.ReadyListener
+
+	// Android only
+	TunAdapter       device.TunAdapter
+	IFaceDiscover    stdnet.ExternalIFaceDiscover
+	HostDNSAddresses []netip.AddrPort
+	DnsReadyListener dns.ReadyListener
 
 	//	iOS only
 	DnsManager     dns.IosDnsManager

client/internal/routemanager/manager.go

@@ -52,6 +52,10 @@ type Manager interface {
 	UpdateRoutes(updateSerial uint64, serverRoutes map[route.ID]*route.Route, clientRoutes route.HAMap, useNewDNSRoute bool) error
 	ClassifyRoutes(newRoutes []*route.Route) (map[route.ID]*route.Route, route.HAMap)
 	TriggerSelection(route.HAMap)
+	SelectRoutes(ids []route.NetID, appendRoute bool) error
+	DeselectRoutes(ids []route.NetID) error
+	SelectAllRoutes()
+	DeselectAllRoutes()
 	GetRouteSelector() *routeselector.RouteSelector
 	GetClientRoutes() route.HAMap
 	GetSelectedClientRoutes() route.HAMap
@@ -800,7 +804,7 @@ func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeI
 	var info exitNodeInfo
 
 	for haID, routes := range clientRoutes {
-		if !m.isExitNodeRoute(routes) {
+		if !isExitNodeRoutes(routes) {
 			continue
 		}
 
@@ -820,13 +824,6 @@ func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeI
 	return info
 }
 
-func (m *DefaultManager) isExitNodeRoute(routes []*route.Route) bool {
-	if len(routes) == 0 {
-		return false
-	}
-	return route.IsV4DefaultRoute(routes[0].Network) || route.IsV6DefaultRoute(routes[0].Network)
-}
-
 func (m *DefaultManager) categorizeUserSelection(netID route.NetID, info *exitNodeInfo) {
 	if m.routeSelector.IsSelected(netID) {
 		info.userSelected = append(info.userSelected, netID)

client/internal/routemanager/mock.go

@@ -16,6 +16,8 @@ type MockManager struct {
 	ClassifyRoutesFunc           func(routes []*route.Route) (map[route.ID]*route.Route, route.HAMap)
 	UpdateRoutesFunc             func(updateSerial uint64, serverRoutes map[route.ID]*route.Route, clientRoutes route.HAMap, useNewDNSRoute bool) error
 	TriggerSelectionFunc         func(haMap route.HAMap)
+	SelectRoutesFunc             func(ids []route.NetID, appendRoute bool) error
+	DeselectRoutesFunc           func(ids []route.NetID) error
 	GetRouteSelectorFunc         func() *routeselector.RouteSelector
 	GetClientRoutesFunc          func() route.HAMap
 	GetSelectedClientRoutesFunc  func() route.HAMap
@@ -55,6 +57,30 @@ func (m *MockManager) TriggerSelection(networks route.HAMap) {
 	}
 }
 
+// SelectRoutes mock implementation of SelectRoutes from Manager interface
+func (m *MockManager) SelectRoutes(ids []route.NetID, appendRoute bool) error {
+	if m.SelectRoutesFunc != nil {
+		return m.SelectRoutesFunc(ids, appendRoute)
+	}
+	return nil
+}
+
+// DeselectRoutes mock implementation of DeselectRoutes from Manager interface
+func (m *MockManager) DeselectRoutes(ids []route.NetID) error {
+	if m.DeselectRoutesFunc != nil {
+		return m.DeselectRoutesFunc(ids)
+	}
+	return nil
+}
+
+// SelectAllRoutes mock implementation of SelectAllRoutes from Manager interface
+func (m *MockManager) SelectAllRoutes() {
+}
+
+// DeselectAllRoutes mock implementation of DeselectAllRoutes from Manager interface
+func (m *MockManager) DeselectAllRoutes() {
+}
+
 // GetRouteSelector mock implementation of GetRouteSelector from Manager interface
 func (m *MockManager) GetRouteSelector() *routeselector.RouteSelector {
 	if m.GetRouteSelectorFunc != nil {

client/internal/routemanager/notifier/notifier_ios.go

@@ -3,7 +3,6 @@
 package notifier
 
 import (
-	"container/list"
 	"net/netip"
 	"slices"
 	"sort"
@@ -16,20 +15,12 @@ import (
 
 type Notifier struct {
 	mu              sync.Mutex
-	cond            *sync.Cond
 	currentPrefixes []string
 	listener        listener.NetworkChangeListener
-	queue           *list.List
-	closed          bool
 }
 
 func NewNotifier() *Notifier {
-	n := &Notifier{
-		queue: list.New(),
-	}
-	n.cond = sync.NewCond(&n.mu)
-	go n.deliverLoop()
-	return n
+	return &Notifier{}
 }
 
 func (n *Notifier) SetListener(listener listener.NetworkChangeListener) {
@@ -59,44 +50,19 @@ func (n *Notifier) OnNewPrefixes(prefixes []netip.Prefix) {
 	sort.Strings(newNets)
 
 	n.mu.Lock()
+	defer n.mu.Unlock()
 	if slices.Equal(n.currentPrefixes, newNets) {
-		n.mu.Unlock()
 		return
 	}
 	n.currentPrefixes = newNets
-	routes := strings.Join(n.currentPrefixes, ",")
-	n.queue.PushBack(routes)
-	n.cond.Signal()
-	n.mu.Unlock()
+	if n.listener != nil {
+		n.listener.OnNetworkChanged(strings.Join(n.currentPrefixes, ","))
+	}
 }
 
 func (n *Notifier) Close() {
-	n.mu.Lock()
-	n.closed = true
-	n.cond.Signal()
-	n.mu.Unlock()
 }
 
 func (n *Notifier) GetInitialRouteRanges() []string {
 	return nil
 }
-
-func (n *Notifier) deliverLoop() {
-	for {
-		n.mu.Lock()
-		for n.queue.Len() == 0 && !n.closed {
-			n.cond.Wait()
-		}
-		if n.closed && n.queue.Len() == 0 {
-			n.mu.Unlock()
-			return
-		}
-		routes := n.queue.Remove(n.queue.Front()).(string)
-		l := n.listener
-		n.mu.Unlock()
-
-		if l != nil {
-			l.OnNetworkChanged(routes)
-		}
-	}
-}

client/internal/routemanager/reconcile_test.go

@@ -54,7 +54,7 @@ func (m *reconcileWGMock) GetNet() *netstack.Net                      { return n
 func TestReconcilePeerAllowedIPs(t *testing.T) {
 	wg := &reconcileWGMock{}
 	m := &DefaultManager{wgInterface: wg}
-	m.allowedIPsRefCounter = refcounter.New[netip.Prefix, string, string](
+	m.allowedIPsRefCounter = refcounter.NewAllowedIPs(
 		func(_ netip.Prefix, peerKey string) (string, error) { return peerKey, nil },
 		func(netip.Prefix, string) error { return nil },
 	)

client/internal/routemanager/refcounter/allowedips.go

@@ -169,6 +169,27 @@ func (rm *AllowedIPsRefCounter) Flush() error {
 	return nberrors.FormatErrorOrNil(merr)
 }
 
+// ReapplyMatching calls apply for every prefix whose currently installed (active) peer satisfies
+// pred, holding the lock for the whole pass. It is used to re-push allowed IPs onto a peer whose
+// WireGuard entry was rebuilt (e.g. a lazy connection cycling idle->wake) without a matching
+// refcounter change, which would otherwise leave the prefix installed in the counter but missing
+// on the device. Only the active peer is considered — a prefix that lost its installed peer to a
+// failed swap is skipped here and reconciled by the next Increment/Decrement.
+func (rm *AllowedIPsRefCounter) ReapplyMatching(pred func(out string) bool, apply func(key netip.Prefix) error) error {
+	rm.mu.Lock()
+	defer rm.mu.Unlock()
+
+	var merr *multierror.Error
+	for prefix, e := range rm.entries {
+		if e.active != "" && pred(e.active) {
+			if err := apply(prefix); err != nil {
+				merr = multierror.Append(merr, err)
+			}
+		}
+	}
+	return nberrors.FormatErrorOrNil(merr)
+}
+
 // pickSurvivor deterministically selects a peer still referencing the prefix. WireGuard cannot do
 // multipath for a single prefix, so any surviving peer is a valid winner; the choice is made stable
 // (lowest peerKey) for predictable behavior and testability.

client/internal/routemanager/selection.go

@@ -0,0 +1,138 @@
+package routemanager
+
+import (
+	"fmt"
+	"slices"
+
+	"github.com/hashicorp/go-multierror"
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/exp/maps"
+
+	nberrors "github.com/netbirdio/netbird/client/errors"
+	"github.com/netbirdio/netbird/route"
+)
+
+// SelectRoutes selects the routes with the given network IDs and applies the
+// new selection. V4/v6 exit-node pairs are expanded automatically. Exit nodes
+// are mutually exclusive: if the selection activates an exit node, every other
+// available exit node is deselected so two can't be active at once. With
+// appendRoute=false the previous selection is replaced instead of extended.
+func (m *DefaultManager) SelectRoutes(ids []route.NetID, appendRoute bool) error {
+	if err := m.selectRoutes(ids, appendRoute); err != nil {
+		return err
+	}
+	m.TriggerSelection(m.GetClientRoutes())
+	return nil
+}
+
+// DeselectRoutes removes the routes with the given network IDs from the
+// selection and applies the change. V4/v6 exit-node pairs are expanded
+// automatically.
+func (m *DefaultManager) DeselectRoutes(ids []route.NetID) error {
+	if err := m.deselectRoutes(ids); err != nil {
+		return err
+	}
+	m.TriggerSelection(m.GetClientRoutes())
+	return nil
+}
+
+func (m *DefaultManager) deselectRoutes(ids []route.NetID) error {
+	routesMap := m.GetClientRoutesWithNetID()
+	routes := route.ExpandV6ExitPairs(slices.Clone(ids), routesMap)
+
+	log.Debugf("deselecting routes with ids: %v", routes)
+
+	if err := m.routeSelector.DeselectRoutes(routes, maps.Keys(routesMap)); err != nil {
+		return fmt.Errorf("deselect routes: %w", err)
+	}
+
+	return nil
+}
+
+// SelectAllRoutes selects every available route and applies the selection.
+// Exit nodes stay mutually exclusive: at most one remains active.
+func (m *DefaultManager) SelectAllRoutes() {
+	m.selectAllRoutes()
+	m.TriggerSelection(m.GetClientRoutes())
+}
+
+func (m *DefaultManager) selectAllRoutes() {
+	m.routeSelector.SelectAllRoutes()
+
+	// Select-all wipes every explicit selection, so exit nodes fall back to
+	// management's auto-apply flags — which may mark several at once.
+	// Reconcile immediately so at most one exit node stays active instead of
+	// waiting for the next network map to enforce it.
+	m.mux.Lock()
+	defer m.mux.Unlock()
+	m.updateRouteSelectorFromManagement(m.clientRoutes)
+}
+
+// DeselectAllRoutes deselects every route and applies the change.
+func (m *DefaultManager) DeselectAllRoutes() {
+	m.routeSelector.DeselectAllRoutes()
+	m.TriggerSelection(m.GetClientRoutes())
+}
+
+func (m *DefaultManager) selectRoutes(ids []route.NetID, appendRoute bool) error {
+	routesMap := m.GetClientRoutesWithNetID()
+	routes := route.ExpandV6ExitPairs(slices.Clone(ids), routesMap)
+	allIDs := maps.Keys(routesMap)
+
+	log.Debugf("selecting routes with ids: %v", routes)
+
+	// A partial failure (e.g. an unknown ID in the request) still selects the
+	// valid routes, so exclusivity below must run regardless of the error.
+	var merr *multierror.Error
+	if err := m.routeSelector.SelectRoutes(routes, appendRoute, allIDs); err != nil {
+		merr = multierror.Append(merr, fmt.Errorf("select routes: %w", err))
+	}
+
+	// Exit nodes are mutually exclusive: if this selection activates an
+	// exit node, deselect every other available exit node so two can't be
+	// selected at once. Non-exit route selections are left untouched.
+	if requestActivatesExitNode(routes, routesMap) {
+		if others := otherExitNodeIDs(routesMap, routes); len(others) > 0 {
+			if err := m.routeSelector.DeselectRoutes(others, allIDs); err != nil {
+				merr = multierror.Append(merr, fmt.Errorf("deselect sibling exit nodes: %w", err))
+			}
+		}
+	}
+
+	return nberrors.FormatErrorOrNil(merr)
+}
+
+func isExitNodeRoutes(routes []*route.Route) bool {
+	return len(routes) > 0 && (route.IsV4DefaultRoute(routes[0].Network) || route.IsV6DefaultRoute(routes[0].Network))
+}
+
+// requestActivatesExitNode reports whether any requested NetID maps to an exit
+// node (default route) in the current route table.
+func requestActivatesExitNode(requested []route.NetID, routesMap map[route.NetID][]*route.Route) bool {
+	for _, id := range requested {
+		if isExitNodeRoutes(routesMap[id]) {
+			return true
+		}
+	}
+	return false
+}
+
+// otherExitNodeIDs returns every available exit-node NetID that is not in the
+// requested set — the siblings to deselect so a single exit node stays active.
+func otherExitNodeIDs(routesMap map[route.NetID][]*route.Route, requested []route.NetID) []route.NetID {
+	keep := make(map[route.NetID]struct{}, len(requested))
+	for _, id := range requested {
+		keep[id] = struct{}{}
+	}
+	var others []route.NetID
+	for id, routes := range routesMap {
+		if !isExitNodeRoutes(routes) {
+			continue
+		}
+		if _, ok := keep[id]; ok {
+			continue
+		}
+		others = append(others, id)
+	}
+	return others
+}

client/internal/routemanager/selection_test.go

@@ -0,0 +1,129 @@
+package routemanager
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/routeselector"
+	"github.com/netbirdio/netbird/route"
+)
+
+func v6ExitRoute(netID, peer string) *route.Route {
+	return &route.Route{
+		NetID:   route.NetID(netID),
+		Network: netip.MustParsePrefix("::/0"),
+		Peer:    peer,
+	}
+}
+
+func newSelectionTestManager() *DefaultManager {
+	return &DefaultManager{
+		routeSelector: routeselector.NewRouteSelector(),
+		clientRoutes: route.HAMap{
+			"exitA|0.0.0.0/0":    {exitRoute("exitA", "p1", true)},
+			"exitA-v6|::/0":      {v6ExitRoute("exitA-v6", "p1")},
+			"exitB|0.0.0.0/0":    {exitRoute("exitB", "p2", true)},
+			"lan|192.168.1.0/24": {{NetID: "lan", Network: netip.MustParsePrefix("192.168.1.0/24"), Peer: "p3"}},
+		},
+	}
+}
+
+func TestSelectRoutes_ExitNodeExclusivity(t *testing.T) {
+	m := newSelectionTestManager()
+
+	// Selecting an exit node selects its v6 pair and deselects the sibling.
+	require.NoError(t, m.selectRoutes([]route.NetID{"exitA"}, true))
+	assert.True(t, m.routeSelector.IsSelected("exitA"), "exitA should be selected")
+	assert.True(t, m.routeSelector.IsSelected("exitA-v6"), "the v6 pair follows its v4 base")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "the sibling exit node must be deselected")
+
+	// Switching to the sibling deselects the previous exit node and its v6 pair.
+	require.NoError(t, m.selectRoutes([]route.NetID{"exitB"}, true))
+	assert.True(t, m.routeSelector.IsSelected("exitB"), "exitB should now be selected")
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "the previous exit node must be deselected")
+	assert.False(t, m.routeSelector.IsSelected("exitA-v6"), "the previous exit node's v6 pair must be deselected")
+	assert.True(t, m.routeSelector.IsSelected("lan"), "non-exit route selection is untouched")
+
+	// Selecting a non-exit route leaves the active exit node alone.
+	require.NoError(t, m.selectRoutes([]route.NetID{"lan"}, true))
+	assert.True(t, m.routeSelector.IsSelected("exitB"), "selecting a non-exit route keeps the exit node")
+
+	// Deselecting the active exit node turns every exit node off.
+	require.NoError(t, m.deselectRoutes([]route.NetID{"exitB"}))
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "exitB should be deselected")
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "exitA stays deselected")
+	assert.True(t, m.routeSelector.IsSelected("lan"), "non-exit route selection is untouched")
+}
+
+func TestSelectRoutes_PartialErrorStillEnforcesExclusivity(t *testing.T) {
+	// The unknown ID must be reported, but the valid exit node in the same
+	// request is still selected — so its sibling must still be deselected.
+	// Both orderings are covered: processing must continue past the invalid
+	// ID wherever it sits in the request.
+	requests := map[string][]route.NetID{
+		"invalid id first": {"missing", "exitB"},
+		"invalid id last":  {"exitB", "missing"},
+	}
+
+	for name, ids := range requests {
+		t.Run(name, func(t *testing.T) {
+			m := newSelectionTestManager()
+
+			require.NoError(t, m.selectRoutes([]route.NetID{"exitA"}, true))
+
+			err := m.selectRoutes(ids, true)
+			assert.Error(t, err, "unknown id must be reported")
+			assert.True(t, m.routeSelector.IsSelected("exitB"), "valid exit node from the request is selected")
+			assert.False(t, m.routeSelector.IsSelected("exitA"), "sibling exit node must be deselected despite the error")
+			assert.False(t, m.routeSelector.IsSelected("exitA-v6"), "sibling's v6 pair must be deselected too")
+		})
+	}
+}
+
+func TestSelectAllRoutes_KeepsSingleExitNode(t *testing.T) {
+	// Both exit nodes are marked for auto-apply by management
+	// (SkipAutoApply=false), the state where select-all could turn on two at
+	// once without the immediate reconciliation.
+	m := &DefaultManager{
+		routeSelector: routeselector.NewRouteSelector(),
+		clientRoutes: route.HAMap{
+			"exitA|0.0.0.0/0":    {exitRoute("exitA", "p1", false)},
+			"exitB|0.0.0.0/0":    {exitRoute("exitB", "p2", false)},
+			"lan|192.168.1.0/24": {{NetID: "lan", Network: netip.MustParsePrefix("192.168.1.0/24"), Peer: "p3"}},
+		},
+	}
+
+	require.NoError(t, m.selectRoutes([]route.NetID{"exitB"}, true))
+
+	m.selectAllRoutes()
+
+	assert.True(t, m.routeSelector.IsSelected("lan"), "non-exit routes are all selected")
+	assert.True(t, m.routeSelector.IsSelected("exitA"), "the deterministic management pick stays active")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "select-all must not leave a second exit node active")
+}
+
+func TestSelectRoutes_UnknownRoute(t *testing.T) {
+	m := newSelectionTestManager()
+
+	assert.Error(t, m.selectRoutes([]route.NetID{"missing"}, true), "selecting an unavailable route must fail")
+	assert.Error(t, m.deselectRoutes([]route.NetID{"missing"}), "deselecting an unavailable route must fail")
+}
+
+func TestExitNodeSelectionHelpers(t *testing.T) {
+	routesMap := map[route.NetID][]*route.Route{
+		"exitA": {{Network: netip.MustParsePrefix("0.0.0.0/0")}},
+		"exitB": {{Network: netip.MustParsePrefix("::/0")}},
+		"lan":   {{Network: netip.MustParsePrefix("192.168.0.0/16")}},
+	}
+
+	assert.True(t, requestActivatesExitNode([]route.NetID{"exitA"}, routesMap), "v4 default route is an exit node")
+	assert.True(t, requestActivatesExitNode([]route.NetID{"exitB"}, routesMap), "v6 default route is an exit node")
+	assert.False(t, requestActivatesExitNode([]route.NetID{"lan"}, routesMap), "lan route is not an exit node")
+	assert.False(t, requestActivatesExitNode([]route.NetID{"missing"}, routesMap), "unknown id is not an exit node")
+
+	others := otherExitNodeIDs(routesMap, []route.NetID{"exitB"})
+	assert.ElementsMatch(t, []route.NetID{"exitA"}, others, "only the other exit node is a sibling; the lan route is ignored")
+}

client/internal/routemanager/sysctl/sysctl_linux.go

@@ -20,6 +20,8 @@ const (
 	rpFilterPath          = "net.ipv4.conf.all.rp_filter"
 	rpFilterInterfacePath = "net.ipv4.conf.%s.rp_filter"
 	srcValidMarkPath      = "net.ipv4.conf.all.src_valid_mark"
+	percentEscape         = "%25"
+	dotEscape             = "%2E"
 )
 
 type iface interface {
@@ -56,7 +58,11 @@ func Setup(wgIface iface) (map[string]int, error) {
 			continue
 		}
 
-		i := fmt.Sprintf(rpFilterInterfacePath, intf.Name)
+		// Escape '%' and '.' so they survive the dot-to-slash conversion in Set()
+		safeName := strings.ReplaceAll(intf.Name, "%", percentEscape)
+		safeName = strings.ReplaceAll(safeName, ".", dotEscape)
+
+		i := fmt.Sprintf(rpFilterInterfacePath, safeName)
 		oldVal, err := Set(i, 2, true)
 		if err != nil {
 			result = multierror.Append(result, err)
@@ -70,7 +76,11 @@ func Setup(wgIface iface) (map[string]int, error) {
 
 // Set sets a sysctl configuration, if onlyIfOne is true it will only set the new value if it's set to 1
 func Set(key string, desiredValue int, onlyIfOne bool) (int, error) {
-	path := fmt.Sprintf("/proc/sys/%s", strings.ReplaceAll(key, ".", "/"))
+	path := strings.ReplaceAll(key, ".", "/")
+	// Unescape interface dots and percent signs
+	path = strings.ReplaceAll(path, dotEscape, ".")
+	path = strings.ReplaceAll(path, percentEscape, "%")
+	path = fmt.Sprintf("/proc/sys/%s", path)
 	currentValue, err := os.ReadFile(path)
 	if err != nil {
 		return -1, fmt.Errorf("read sysctl %s: %w", key, err)

client/internal/tunnelnotifier/notifier.go

@@ -0,0 +1,124 @@
+package tunnelnotifier
+
+import (
+	"container/list"
+	"sync"
+
+	"github.com/netbirdio/netbird/client/internal/dns"
+	"github.com/netbirdio/netbird/client/internal/listener"
+)
+
+type eventKind int
+
+const (
+	eventRoutes eventKind = iota
+	eventIfaceIP
+	eventIfaceIPv6
+	eventDNS
+)
+
+var (
+	_ listener.NetworkChangeListener = (*Notifier)(nil)
+	_ dns.IosDnsManager              = (*Notifier)(nil)
+)
+
+type event struct {
+	kind    eventKind
+	payload string
+}
+
+type Notifier struct {
+	mu     sync.Mutex
+	cond   *sync.Cond
+	queue  *list.List
+	closed bool
+	done   chan struct{}
+
+	listener   listener.NetworkChangeListener
+	dnsManager dns.IosDnsManager
+}
+
+func New(l listener.NetworkChangeListener, dm dns.IosDnsManager) *Notifier {
+	n := &Notifier{
+		queue:      list.New(),
+		done:       make(chan struct{}),
+		listener:   l,
+		dnsManager: dm,
+	}
+	n.cond = sync.NewCond(&n.mu)
+	go n.deliverLoop()
+	return n
+}
+
+func (n *Notifier) OnNetworkChanged(routes string) {
+	n.enqueue(event{kind: eventRoutes, payload: routes})
+}
+
+func (n *Notifier) SetInterfaceIP(ip string) {
+	n.enqueue(event{kind: eventIfaceIP, payload: ip})
+}
+
+func (n *Notifier) SetInterfaceIPv6(ip string) {
+	n.enqueue(event{kind: eventIfaceIPv6, payload: ip})
+}
+
+func (n *Notifier) ApplyDns(config string) {
+	n.enqueue(event{kind: eventDNS, payload: config})
+}
+
+// Close stops accepting new events and blocks until the delivery loop has
+// drained all queued events and exited.
+func (n *Notifier) Close() {
+	n.mu.Lock()
+	n.closed = true
+	n.cond.Signal()
+	n.mu.Unlock()
+	<-n.done
+}
+
+func (n *Notifier) enqueue(ev event) {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	if n.closed {
+		return
+	}
+	n.queue.PushBack(ev)
+	n.cond.Signal()
+}
+
+func (n *Notifier) deliverLoop() {
+	defer close(n.done)
+	for {
+		n.mu.Lock()
+		for n.queue.Len() == 0 && !n.closed {
+			n.cond.Wait()
+		}
+		if n.closed && n.queue.Len() == 0 {
+			n.mu.Unlock()
+			return
+		}
+		ev := n.queue.Remove(n.queue.Front()).(event)
+		l := n.listener
+		dm := n.dnsManager
+		n.mu.Unlock()
+
+		switch ev.kind {
+		case eventRoutes:
+			if l != nil {
+				l.OnNetworkChanged(ev.payload)
+			}
+		case eventIfaceIP:
+			if l != nil {
+				l.SetInterfaceIP(ev.payload)
+			}
+		case eventIfaceIPv6:
+			if l != nil {
+				l.SetInterfaceIPv6(ev.payload)
+			}
+		case eventDNS:
+			if dm != nil {
+				dm.ApplyDns(ev.payload)
+			}
+		}
+	}
+}

client/internal/tunnelnotifier/notifier_test.go

@@ -0,0 +1,192 @@
+package tunnelnotifier
+
+import (
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+type call struct {
+	kind    string
+	payload string
+}
+
+type recorder struct {
+	mu       sync.Mutex
+	calls    []call
+	inFlight atomic.Int32
+	overlap  atomic.Bool
+	delay    time.Duration
+}
+
+func (r *recorder) record(kind, payload string) {
+	if r.inFlight.Add(1) != 1 {
+		r.overlap.Store(true)
+	}
+	if r.delay > 0 {
+		time.Sleep(r.delay)
+	}
+	r.mu.Lock()
+	r.calls = append(r.calls, call{kind: kind, payload: payload})
+	r.mu.Unlock()
+	r.inFlight.Add(-1)
+}
+
+func (r *recorder) count() int {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return len(r.calls)
+}
+
+func (r *recorder) snapshot() []call {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	out := make([]call, len(r.calls))
+	copy(out, r.calls)
+	return out
+}
+
+type fakeListener struct {
+	rec *recorder
+}
+
+func (f *fakeListener) OnNetworkChanged(routes string) {
+	f.rec.record("routes", routes)
+}
+
+func (f *fakeListener) SetInterfaceIP(ip string) {
+	f.rec.record("ip", ip)
+}
+
+func (f *fakeListener) SetInterfaceIPv6(ip string) {
+	f.rec.record("ipv6", ip)
+}
+
+type fakeDNSManager struct {
+	rec *recorder
+}
+
+func (f *fakeDNSManager) ApplyDns(config string) {
+	f.rec.record("dns", config)
+}
+
+func TestFIFOOrder(t *testing.T) {
+	rec := &recorder{}
+	n := New(&fakeListener{rec: rec}, &fakeDNSManager{rec: rec})
+	defer n.Close()
+
+	n.SetInterfaceIP("10.0.0.1")
+	n.SetInterfaceIPv6("fd00::1")
+	n.ApplyDns(`{"domains":[]}`)
+	n.OnNetworkChanged("10.0.0.0/8,192.168.0.0/16")
+	n.ApplyDns(`{"domains":["example.com"]}`)
+
+	require.Eventually(t, func() bool { return rec.count() == 5 }, time.Second, time.Millisecond)
+
+	expected := []call{
+		{kind: "ip", payload: "10.0.0.1"},
+		{kind: "ipv6", payload: "fd00::1"},
+		{kind: "dns", payload: `{"domains":[]}`},
+		{kind: "routes", payload: "10.0.0.0/8,192.168.0.0/16"},
+		{kind: "dns", payload: `{"domains":["example.com"]}`},
+	}
+	assert.Equal(t, expected, rec.snapshot())
+}
+
+func TestNoOverlappingCalls(t *testing.T) {
+	rec := &recorder{delay: 100 * time.Microsecond}
+	n := New(&fakeListener{rec: rec}, &fakeDNSManager{rec: rec})
+	defer n.Close()
+
+	const producers = 8
+	const perProducer = 25
+
+	var wg sync.WaitGroup
+	for i := 0; i < producers; i++ {
+		wg.Add(1)
+		go func(id int) {
+			defer wg.Done()
+			for j := 0; j < perProducer; j++ {
+				payload := fmt.Sprintf("%d-%d", id, j)
+				switch j % 4 {
+				case 0:
+					n.OnNetworkChanged(payload)
+				case 1:
+					n.SetInterfaceIP(payload)
+				case 2:
+					n.SetInterfaceIPv6(payload)
+				case 3:
+					n.ApplyDns(payload)
+				}
+			}
+		}(i)
+	}
+	wg.Wait()
+
+	require.Eventually(t, func() bool { return rec.count() == producers*perProducer }, 5*time.Second, time.Millisecond)
+	assert.False(t, rec.overlap.Load())
+}
+
+func TestDNSAndRoutesInterleaved(t *testing.T) {
+	rec := &recorder{delay: 100 * time.Microsecond}
+	n := New(&fakeListener{rec: rec}, &fakeDNSManager{rec: rec})
+	defer n.Close()
+
+	const events = 50
+
+	var wg sync.WaitGroup
+	wg.Add(2)
+	go func() {
+		defer wg.Done()
+		for i := 0; i < events; i++ {
+			n.ApplyDns(fmt.Sprintf("dns-%d", i))
+		}
+	}()
+	go func() {
+		defer wg.Done()
+		for i := 0; i < events; i++ {
+			n.OnNetworkChanged(fmt.Sprintf("routes-%d", i))
+		}
+	}()
+	wg.Wait()
+
+	require.Eventually(t, func() bool { return rec.count() == 2*events }, 5*time.Second, time.Millisecond)
+	assert.False(t, rec.overlap.Load())
+
+	var dnsSeen, routesSeen int
+	for _, c := range rec.snapshot() {
+		switch c.kind {
+		case "dns":
+			assert.Equal(t, fmt.Sprintf("dns-%d", dnsSeen), c.payload)
+			dnsSeen++
+		case "routes":
+			assert.Equal(t, fmt.Sprintf("routes-%d", routesSeen), c.payload)
+			routesSeen++
+		}
+	}
+	assert.Equal(t, events, dnsSeen)
+	assert.Equal(t, events, routesSeen)
+}
+
+func TestCloseDrainsQueue(t *testing.T) {
+	rec := &recorder{delay: time.Millisecond}
+	n := New(&fakeListener{rec: rec}, &fakeDNSManager{rec: rec})
+
+	const events = 20
+	for i := 0; i < events; i++ {
+		n.OnNetworkChanged(fmt.Sprintf("routes-%d", i))
+	}
+	n.Close()
+
+	require.Equal(t, events, rec.count(), "Close must not return before all queued events are delivered")
+
+	n.OnNetworkChanged("after-close")
+	n.ApplyDns("after-close")
+	time.Sleep(50 * time.Millisecond)
+	assert.Equal(t, events, rec.count())
+}

client/ios/NetBirdSDK/client.go

@@ -13,7 +13,6 @@ import (
 	"time"
 
 	log "github.com/sirupsen/logrus"
-	"golang.org/x/exp/maps"
 
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/auth"
@@ -637,23 +636,18 @@ func (c *Client) SelectRoute(id string) error {
 	}
 
 	routeManager := engine.GetRouteManager()
-	routeSelector := routeManager.GetRouteSelector()
 	if id == "All" {
 		log.Debugf("select all routes")
-		routeSelector.SelectAllRoutes()
-	} else {
-		log.Debugf("select route with id: %s", id)
-		routes := toNetIDs([]string{id})
-		routesMap := routeManager.GetClientRoutesWithNetID()
-		routes = route.ExpandV6ExitPairs(routes, routesMap)
-		if err := routeSelector.SelectRoutes(routes, true, maps.Keys(routesMap)); err != nil {
-			log.Debugf("error when selecting routes: %s", err)
-			return fmt.Errorf("select routes: %w", err)
-		}
+		routeManager.SelectAllRoutes()
+		return nil
 	}
-	routeManager.TriggerSelection(routeManager.GetClientRoutes())
-	return nil
 
+	log.Debugf("select route with id: %s", id)
+	if err := routeManager.SelectRoutes(toNetIDs([]string{id}), true); err != nil {
+		log.Debugf("error when selecting routes: %s", err)
+		return err
+	}
+	return nil
 }
 
 func (c *Client) DeselectRoute(id string) error {
@@ -667,21 +661,17 @@ func (c *Client) DeselectRoute(id string) error {
 	}
 
 	routeManager := engine.GetRouteManager()
-	routeSelector := routeManager.GetRouteSelector()
 	if id == "All" {
 		log.Debugf("deselect all routes")
-		routeSelector.DeselectAllRoutes()
-	} else {
-		log.Debugf("deselect route with id: %s", id)
-		routes := toNetIDs([]string{id})
-		routesMap := routeManager.GetClientRoutesWithNetID()
-		routes = route.ExpandV6ExitPairs(routes, routesMap)
-		if err := routeSelector.DeselectRoutes(routes, maps.Keys(routesMap)); err != nil {
-			log.Debugf("error when deselecting routes: %s", err)
-			return fmt.Errorf("deselect routes: %w", err)
-		}
+		routeManager.DeselectAllRoutes()
+		return nil
+	}
+
+	log.Debugf("deselect route with id: %s", id)
+	if err := routeManager.DeselectRoutes(toNetIDs([]string{id})); err != nil {
+		log.Debugf("error when deselecting routes: %s", err)
+		return err
 	}
-	routeManager.TriggerSelection(routeManager.GetClientRoutes())
 	return nil
 }
 

client/server/network.go

@@ -8,7 +8,6 @@ import (
 	"sort"
 	"strings"
 
-	"golang.org/x/exp/maps"
 	"google.golang.org/grpc/codes"
 	gstatus "google.golang.org/grpc/status"
 
@@ -161,30 +160,11 @@ func (s *Server) SelectNetworks(_ context.Context, req *proto.SelectNetworksRequ
 		return nil, fmt.Errorf("no route manager")
 	}
 
-	routeSelector := routeManager.GetRouteSelector()
 	if req.GetAll() {
-		routeSelector.SelectAllRoutes()
-	} else {
-		routes := toNetIDs(req.GetNetworkIDs())
-		routesMap := routeManager.GetClientRoutesWithNetID()
-		routes = route.ExpandV6ExitPairs(routes, routesMap)
-		netIdRoutes := maps.Keys(routesMap)
-		if err := routeSelector.SelectRoutes(routes, req.GetAppend(), netIdRoutes); err != nil {
-			return nil, fmt.Errorf("select routes: %w", err)
-		}
-
-		// Exit nodes are mutually exclusive: if this selection activates an
-		// exit node, deselect every other available exit node so two can't be
-		// selected at once. Non-exit route selections are left untouched.
-		if requestActivatesExitNode(routes, routesMap) {
-			if others := otherExitNodeIDs(routesMap, routes); len(others) > 0 {
-				if err := routeSelector.DeselectRoutes(others, netIdRoutes); err != nil {
-					return nil, fmt.Errorf("deselect sibling exit nodes: %w", err)
-				}
-			}
-		}
+		routeManager.SelectAllRoutes()
+	} else if err := routeManager.SelectRoutes(toNetIDs(req.GetNetworkIDs()), req.GetAppend()); err != nil {
+		return nil, err
 	}
-	routeManager.TriggerSelection(routeManager.GetClientRoutes())
 
 	s.statusRecorder.PublishEvent(
 		proto.SystemEvent_INFO,
@@ -224,19 +204,11 @@ func (s *Server) DeselectNetworks(_ context.Context, req *proto.SelectNetworksRe
 		return nil, fmt.Errorf("no route manager")
 	}
 
-	routeSelector := routeManager.GetRouteSelector()
 	if req.GetAll() {
-		routeSelector.DeselectAllRoutes()
-	} else {
-		routes := toNetIDs(req.GetNetworkIDs())
-		routesMap := routeManager.GetClientRoutesWithNetID()
-		routes = route.ExpandV6ExitPairs(routes, routesMap)
-		netIdRoutes := maps.Keys(routesMap)
-		if err := routeSelector.DeselectRoutes(routes, netIdRoutes); err != nil {
-			return nil, fmt.Errorf("deselect routes: %w", err)
-		}
+		routeManager.DeselectAllRoutes()
+	} else if err := routeManager.DeselectRoutes(toNetIDs(req.GetNetworkIDs())); err != nil {
+		return nil, err
 	}
-	routeManager.TriggerSelection(routeManager.GetClientRoutes())
 
 	s.statusRecorder.PublishEvent(
 		proto.SystemEvent_INFO,
@@ -261,37 +233,3 @@ func toNetIDs(routes []string) []route.NetID {
 	return netIDs
 }
 
-func isExitNodeRoutes(routes []*route.Route) bool {
-	return len(routes) > 0 && (route.IsV4DefaultRoute(routes[0].Network) || route.IsV6DefaultRoute(routes[0].Network))
-}
-
-// requestActivatesExitNode reports whether any requested NetID maps to an exit
-// node (default route) in the current route table.
-func requestActivatesExitNode(requested []route.NetID, routesMap map[route.NetID][]*route.Route) bool {
-	for _, id := range requested {
-		if isExitNodeRoutes(routesMap[id]) {
-			return true
-		}
-	}
-	return false
-}
-
-// otherExitNodeIDs returns every available exit-node NetID that is not in the
-// requested set — the siblings to deselect so a single exit node stays active.
-func otherExitNodeIDs(routesMap map[route.NetID][]*route.Route, requested []route.NetID) []route.NetID {
-	keep := make(map[route.NetID]struct{}, len(requested))
-	for _, id := range requested {
-		keep[id] = struct{}{}
-	}
-	var others []route.NetID
-	for id, routes := range routesMap {
-		if !isExitNodeRoutes(routes) {
-			continue
-		}
-		if _, ok := keep[id]; ok {
-			continue
-		}
-		others = append(others, id)
-	}
-	return others
-}

client/server/network_exitnode_test.go

@@ -1,26 +0,0 @@
-package server
-
-import (
-	"net/netip"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-
-	"github.com/netbirdio/netbird/route"
-)
-
-func TestExitNodeSelectionHelpers(t *testing.T) {
-	routesMap := map[route.NetID][]*route.Route{
-		"exitA": {{Network: netip.MustParsePrefix("0.0.0.0/0")}},
-		"exitB": {{Network: netip.MustParsePrefix("::/0")}},
-		"lan":   {{Network: netip.MustParsePrefix("192.168.0.0/16")}},
-	}
-
-	assert.True(t, requestActivatesExitNode([]route.NetID{"exitA"}, routesMap), "v4 default route is an exit node")
-	assert.True(t, requestActivatesExitNode([]route.NetID{"exitB"}, routesMap), "v6 default route is an exit node")
-	assert.False(t, requestActivatesExitNode([]route.NetID{"lan"}, routesMap), "lan route is not an exit node")
-	assert.False(t, requestActivatesExitNode([]route.NetID{"missing"}, routesMap), "unknown id is not an exit node")
-
-	others := otherExitNodeIDs(routesMap, []route.NetID{"exitB"})
-	assert.ElementsMatch(t, []route.NetID{"exitA"}, others, "only the other exit node is a sibling; the lan route is ignored")
-}

management/cmd/management.go

@@ -23,6 +23,7 @@ import (
 	"github.com/netbirdio/netbird/management/server/types"
 
 	"github.com/netbirdio/netbird/formatter/hook"
+	agentnetworkpricing "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
 	"github.com/netbirdio/netbird/management/internals/server"
 	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
 	nbdomain "github.com/netbirdio/netbird/shared/management/domain"
@@ -112,6 +113,23 @@ var (
 				mgmtSingleAccModeDomain = ""
 			}
 
+			// Load the management-side LLM pricing defaults file: an
+			// explicitly configured path is required to load (a typo must
+			// fail startup — the operator believes those rates are live);
+			// otherwise <datadir>/defaults_llm_pricing.yaml is probed and
+			// may be absent (compiled-in defaults serve). Either way the
+			// path stays watched: the reloader picks up edits — and the
+			// file appearing later — without a restart.
+			pricingPath := config.AgentNetwork.PricingDefaultsFile
+			pricingRequired := pricingPath != ""
+			if !pricingRequired {
+				pricingPath = filepath.Join(config.Datadir, agentnetworkpricing.DefaultFileName)
+			}
+			if err := agentnetworkpricing.LoadFile(pricingPath, pricingRequired); err != nil {
+				return fmt.Errorf("load agent-network pricing defaults: %v", err)
+			}
+			agentnetworkpricing.StartReloader(ctx, agentnetworkpricing.ReloadInterval)
+
 			srv := newServer(&server.Config{
 				NbConfig:                    config,
 				DNSDomain:                   dnsDomain,

management/internals/modules/agentnetwork/catalog/catalog.go

@@ -7,12 +7,28 @@ package catalog
 import "github.com/netbirdio/netbird/shared/management/http/api"
 
 // Model is the in-memory representation of a catalog model.
+//
+// The three cache rates mirror the proxy cost meter's Entry semantics
+// (USD per 1k tokens; 0 = no rate configured, that bucket bills at
+// InputPer1k):
+//   - CachedInputPer1k: OpenAI-shape rate for cached prompt tokens
+//     (a SUBSET of input tokens). Typically 0.1-0.5x input.
+//   - CacheReadPer1k / CacheCreationPer1k: Anthropic-shape rates for
+//     the two ADDITIVE prompt-cache buckets. Typically 0.1x / 1.25x
+//     input.
+//
+// The catalog is the single default-pricing source: the agentnetwork
+// pricing package folds these models into per-surface tables that the
+// synthesizer ships to the proxy's cost_meter.
 type Model struct {
-	ID            string
-	Label         string
-	InputPer1k    float64
-	OutputPer1k   float64
-	ContextWindow int
+	ID                 string
+	Label              string
+	InputPer1k         float64
+	OutputPer1k        float64
+	CachedInputPer1k   float64
+	CacheReadPer1k     float64
+	CacheCreationPer1k float64
+	ContextWindow      int
 }
 
 // ProviderKind groups catalog entries for UI presentation. The split
@@ -65,6 +81,17 @@ type Provider struct {
 	// surface — the proxy middleware then falls back to URL sniffing
 	// or skips request-side enrichment.
 	ParserID string
+	// PricingSurfaces names the cost-meter pricing surfaces this
+	// provider's Models are priced under ("openai", "anthropic",
+	// "bedrock" — the llm.Parser surface the request parser stamps as
+	// llm.provider at billing time). NOT derivable from ParserID:
+	// bedrock_api and vertex_ai_api leave ParserID empty (URL-sniffed)
+	// yet price under "bedrock" / "anthropic", and kimi_api serves two
+	// body shapes so it prices under both. Nil for gateway/custom
+	// entries, which declare no models. Same (surface, model) pair
+	// contributed by two providers must carry identical rates — the
+	// pricing package's tests enforce that.
+	PricingSurfaces []string
 	// IdentityInjection, when non-nil, instructs the proxy to stamp
 	// the caller's NetBird identity onto upstream requests under the
 	// configured header names. Used for gateways like LiteLLM that
@@ -219,6 +246,7 @@ var providers = []Provider{
 		DefaultContentType: "application/json",
 		BrandColor:         "#10A37F",
 		ParserID:           "openai",
+		PricingSurfaces:    []string{"openai"},
 		// Pricing + context windows cross-checked against LiteLLM's
 		// model_prices_and_context_window.json. Notable corrections from
 		// earlier values: o4-mini repriced from $4/$16 to $1.10/$4.40
@@ -226,20 +254,20 @@ var providers = []Provider{
 		// family context windows split between 1.05M for full-size
 		// models and 272K for mini/nano/codex variants.
 		Models: []Model{
-			{ID: "gpt-5.5", Label: "GPT-5.5", InputPer1k: 0.005, OutputPer1k: 0.030, ContextWindow: 1050000},
-			{ID: "gpt-5.5-pro", Label: "GPT-5.5 Pro", InputPer1k: 0.030, OutputPer1k: 0.180, ContextWindow: 1050000},
-			{ID: "gpt-5.4", Label: "GPT-5.4", InputPer1k: 0.0025, OutputPer1k: 0.015, ContextWindow: 1050000},
-			{ID: "gpt-5.4-pro", Label: "GPT-5.4 Pro", InputPer1k: 0.030, OutputPer1k: 0.180, ContextWindow: 1050000},
-			{ID: "gpt-5.4-mini", Label: "GPT-5.4 Mini", InputPer1k: 0.00075, OutputPer1k: 0.0045, ContextWindow: 272000},
-			{ID: "gpt-5.4-nano", Label: "GPT-5.4 Nano", InputPer1k: 0.0002, OutputPer1k: 0.00125, ContextWindow: 272000},
-			{ID: "gpt-5.3-codex", Label: "GPT-5.3 Codex", InputPer1k: 0.00175, OutputPer1k: 0.014, ContextWindow: 272000},
-			{ID: "gpt-5.3-chat-latest", Label: "GPT-5.3 Chat", InputPer1k: 0.00175, OutputPer1k: 0.014, ContextWindow: 128000},
-			{ID: "o4-mini", Label: "o4-mini", InputPer1k: 0.0011, OutputPer1k: 0.0044, ContextWindow: 200000},
-			{ID: "gpt-4.1", Label: "GPT-4.1", InputPer1k: 0.002, OutputPer1k: 0.008, ContextWindow: 1047576},
-			{ID: "gpt-4.1-mini", Label: "GPT-4.1 mini", InputPer1k: 0.0004, OutputPer1k: 0.0016, ContextWindow: 1047576},
-			{ID: "gpt-4.1-nano", Label: "GPT-4.1 nano", InputPer1k: 0.0001, OutputPer1k: 0.0004, ContextWindow: 1047576},
-			{ID: "gpt-4o", Label: "GPT-4o", InputPer1k: 0.0025, OutputPer1k: 0.010, ContextWindow: 128000},
-			{ID: "gpt-4o-mini", Label: "GPT-4o mini", InputPer1k: 0.00015, OutputPer1k: 0.0006, ContextWindow: 128000},
+			{ID: "gpt-5.5", Label: "GPT-5.5", InputPer1k: 0.005, OutputPer1k: 0.030, CachedInputPer1k: 0.0005, ContextWindow: 1050000},
+			{ID: "gpt-5.5-pro", Label: "GPT-5.5 Pro", InputPer1k: 0.030, OutputPer1k: 0.180, CachedInputPer1k: 0.003, ContextWindow: 1050000},
+			{ID: "gpt-5.4", Label: "GPT-5.4", InputPer1k: 0.0025, OutputPer1k: 0.015, CachedInputPer1k: 0.00025, ContextWindow: 1050000},
+			{ID: "gpt-5.4-pro", Label: "GPT-5.4 Pro", InputPer1k: 0.030, OutputPer1k: 0.180, CachedInputPer1k: 0.003, ContextWindow: 1050000},
+			{ID: "gpt-5.4-mini", Label: "GPT-5.4 Mini", InputPer1k: 0.00075, OutputPer1k: 0.0045, CachedInputPer1k: 0.000075, ContextWindow: 272000},
+			{ID: "gpt-5.4-nano", Label: "GPT-5.4 Nano", InputPer1k: 0.0002, OutputPer1k: 0.00125, CachedInputPer1k: 0.00002, ContextWindow: 272000},
+			{ID: "gpt-5.3-codex", Label: "GPT-5.3 Codex", InputPer1k: 0.00175, OutputPer1k: 0.014, CachedInputPer1k: 0.000175, ContextWindow: 272000},
+			{ID: "gpt-5.3-chat-latest", Label: "GPT-5.3 Chat", InputPer1k: 0.00175, OutputPer1k: 0.014, CachedInputPer1k: 0.000175, ContextWindow: 128000},
+			{ID: "o4-mini", Label: "o4-mini", InputPer1k: 0.0011, OutputPer1k: 0.0044, CachedInputPer1k: 0.000275, ContextWindow: 200000},
+			{ID: "gpt-4.1", Label: "GPT-4.1", InputPer1k: 0.002, OutputPer1k: 0.008, CachedInputPer1k: 0.0005, ContextWindow: 1047576},
+			{ID: "gpt-4.1-mini", Label: "GPT-4.1 mini", InputPer1k: 0.0004, OutputPer1k: 0.0016, CachedInputPer1k: 0.0001, ContextWindow: 1047576},
+			{ID: "gpt-4.1-nano", Label: "GPT-4.1 nano", InputPer1k: 0.0001, OutputPer1k: 0.0004, CachedInputPer1k: 0.000025, ContextWindow: 1047576},
+			{ID: "gpt-4o", Label: "GPT-4o", InputPer1k: 0.0025, OutputPer1k: 0.010, CachedInputPer1k: 0.00125, ContextWindow: 128000},
+			{ID: "gpt-4o-mini", Label: "GPT-4o mini", InputPer1k: 0.00015, OutputPer1k: 0.0006, CachedInputPer1k: 0.000075, ContextWindow: 128000},
 			{ID: "gpt-4-turbo", Label: "GPT-4 Turbo", InputPer1k: 0.01, OutputPer1k: 0.03, ContextWindow: 128000},
 			{ID: "gpt-3.5-turbo", Label: "GPT-3.5 Turbo", InputPer1k: 0.0005, OutputPer1k: 0.0015, ContextWindow: 16385},
 			{ID: "text-embedding-3-large", Label: "text-embedding-3-large", InputPer1k: 0.00013, OutputPer1k: 0, ContextWindow: 8191},
@@ -257,6 +285,7 @@ var providers = []Provider{
 		DefaultContentType: "application/json",
 		BrandColor:         "#D97757",
 		ParserID:           "anthropic",
+		PricingSurfaces:    []string{"anthropic"},
 		// Per Anthropic's current model lineup. Pricing in USD per 1k
 		// tokens. Context windows: 4.6+ family is 1M; Haiku 4.5 stays at
 		// 200K. claude-3-7-sonnet and claude-3-5-haiku retired
@@ -267,14 +296,14 @@ var providers = []Provider{
 		// account to be on >= 30-day data retention or all requests
 		// 400.
 		Models: []Model{
-			{ID: "claude-fable-5", Label: "Claude Fable 5", InputPer1k: 0.010, OutputPer1k: 0.050, ContextWindow: 1000000},
-			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "claude-opus-4-6", Label: "Claude Opus 4.6", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "claude-opus-4-1", Label: "Claude Opus 4.1 (deprecated, retires 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, ContextWindow: 200000},
-			{ID: "claude-sonnet-4-6", Label: "Claude Sonnet 4.6", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
-			{ID: "claude-sonnet-4-5", Label: "Claude Sonnet 4.5", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 200000},
-			{ID: "claude-haiku-4-5", Label: "Claude Haiku 4.5", InputPer1k: 0.001, OutputPer1k: 0.005, ContextWindow: 200000},
+			{ID: "claude-fable-5", Label: "Claude Fable 5", InputPer1k: 0.010, OutputPer1k: 0.050, CacheReadPer1k: 0.001, CacheCreationPer1k: 0.0125, ContextWindow: 1000000},
+			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "claude-opus-4-6", Label: "Claude Opus 4.6", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "claude-opus-4-1", Label: "Claude Opus 4.1 (deprecated, retires 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, CacheReadPer1k: 0.0015, CacheCreationPer1k: 0.01875, ContextWindow: 200000},
+			{ID: "claude-sonnet-4-6", Label: "Claude Sonnet 4.6", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
+			{ID: "claude-sonnet-4-5", Label: "Claude Sonnet 4.5", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 200000},
+			{ID: "claude-haiku-4-5", Label: "Claude Haiku 4.5", InputPer1k: 0.001, OutputPer1k: 0.005, CacheReadPer1k: 0.0001, CacheCreationPer1k: 0.00125, ContextWindow: 200000},
 		},
 	},
 	{
@@ -288,18 +317,19 @@ var providers = []Provider{
 		DefaultContentType: "application/json",
 		BrandColor:         "#0078D4",
 		ParserID:           "openai",
+		PricingSurfaces:    []string{"openai"},
 		// Mirrors openai_api pricing — Azure resells OpenAI models at the
 		// same per-token rates, just under different deployment names.
 		Models: []Model{
-			{ID: "gpt-5.5", Label: "GPT-5.5 (Azure)", InputPer1k: 0.005, OutputPer1k: 0.030, ContextWindow: 1050000},
-			{ID: "gpt-5.4", Label: "GPT-5.4 (Azure)", InputPer1k: 0.0025, OutputPer1k: 0.015, ContextWindow: 1050000},
-			{ID: "gpt-5.4-mini", Label: "GPT-5.4 Mini (Azure)", InputPer1k: 0.00075, OutputPer1k: 0.0045, ContextWindow: 272000},
-			{ID: "gpt-5.4-nano", Label: "GPT-5.4 Nano (Azure)", InputPer1k: 0.0002, OutputPer1k: 0.00125, ContextWindow: 272000},
-			{ID: "o4-mini", Label: "o4-mini (Azure)", InputPer1k: 0.0011, OutputPer1k: 0.0044, ContextWindow: 200000},
-			{ID: "gpt-4.1", Label: "GPT-4.1 (Azure)", InputPer1k: 0.002, OutputPer1k: 0.008, ContextWindow: 1047576},
-			{ID: "gpt-4.1-mini", Label: "GPT-4.1 mini (Azure)", InputPer1k: 0.0004, OutputPer1k: 0.0016, ContextWindow: 1047576},
-			{ID: "gpt-4o", Label: "GPT-4o (Azure)", InputPer1k: 0.0025, OutputPer1k: 0.010, ContextWindow: 128000},
-			{ID: "gpt-4o-mini", Label: "GPT-4o mini (Azure)", InputPer1k: 0.00015, OutputPer1k: 0.0006, ContextWindow: 128000},
+			{ID: "gpt-5.5", Label: "GPT-5.5 (Azure)", InputPer1k: 0.005, OutputPer1k: 0.030, CachedInputPer1k: 0.0005, ContextWindow: 1050000},
+			{ID: "gpt-5.4", Label: "GPT-5.4 (Azure)", InputPer1k: 0.0025, OutputPer1k: 0.015, CachedInputPer1k: 0.00025, ContextWindow: 1050000},
+			{ID: "gpt-5.4-mini", Label: "GPT-5.4 Mini (Azure)", InputPer1k: 0.00075, OutputPer1k: 0.0045, CachedInputPer1k: 0.000075, ContextWindow: 272000},
+			{ID: "gpt-5.4-nano", Label: "GPT-5.4 Nano (Azure)", InputPer1k: 0.0002, OutputPer1k: 0.00125, CachedInputPer1k: 0.00002, ContextWindow: 272000},
+			{ID: "o4-mini", Label: "o4-mini (Azure)", InputPer1k: 0.0011, OutputPer1k: 0.0044, CachedInputPer1k: 0.000275, ContextWindow: 200000},
+			{ID: "gpt-4.1", Label: "GPT-4.1 (Azure)", InputPer1k: 0.002, OutputPer1k: 0.008, CachedInputPer1k: 0.0005, ContextWindow: 1047576},
+			{ID: "gpt-4.1-mini", Label: "GPT-4.1 mini (Azure)", InputPer1k: 0.0004, OutputPer1k: 0.0016, CachedInputPer1k: 0.0001, ContextWindow: 1047576},
+			{ID: "gpt-4o", Label: "GPT-4o (Azure)", InputPer1k: 0.0025, OutputPer1k: 0.010, CachedInputPer1k: 0.00125, ContextWindow: 128000},
+			{ID: "gpt-4o-mini", Label: "GPT-4o mini (Azure)", InputPer1k: 0.00015, OutputPer1k: 0.0006, CachedInputPer1k: 0.000075, ContextWindow: 128000},
 			{ID: "gpt-35-turbo", Label: "GPT-3.5 Turbo (Azure)", InputPer1k: 0.0005, OutputPer1k: 0.0015, ContextWindow: 16385},
 		},
 	},
@@ -313,6 +343,9 @@ var providers = []Provider{
 		AuthHeaderTemplate: "Bearer ${API_KEY}",
 		DefaultContentType: "application/json",
 		BrandColor:         "#FF9900",
+		// ParserID stays empty (path-style dispatch via IsBedrockPathStyle);
+		// the request parser meters these under the "bedrock" surface.
+		PricingSurfaces: []string{"bedrock"},
 		// Anthropic models on Bedrock take the anthropic.* prefix and
 		// follow the same lineup / pricing as the first-party Anthropic
 		// catalog entry above. claude-3-7-sonnet and claude-3-5-haiku
@@ -322,13 +355,13 @@ var providers = []Provider{
 		// Llama 3.3 70B entry kept unchanged — LiteLLM tracks only
 		// per-region Llama 3 entries; standalone 3.3 not yet listed.
 		Models: []Model{
-			{ID: "anthropic.claude-opus-4-8", Label: "Claude Opus 4.8 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "anthropic.claude-opus-4-7", Label: "Claude Opus 4.7 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "anthropic.claude-opus-4-6", Label: "Claude Opus 4.6 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "anthropic.claude-opus-4-1", Label: "Claude Opus 4.1 (Bedrock, deprecated 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, ContextWindow: 200000},
-			{ID: "anthropic.claude-sonnet-4-6", Label: "Claude Sonnet 4.6 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
-			{ID: "anthropic.claude-sonnet-4-5", Label: "Claude Sonnet 4.5 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 200000},
-			{ID: "anthropic.claude-haiku-4-5", Label: "Claude Haiku 4.5 (Bedrock)", InputPer1k: 0.001, OutputPer1k: 0.005, ContextWindow: 200000},
+			{ID: "anthropic.claude-opus-4-8", Label: "Claude Opus 4.8 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "anthropic.claude-opus-4-7", Label: "Claude Opus 4.7 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "anthropic.claude-opus-4-6", Label: "Claude Opus 4.6 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "anthropic.claude-opus-4-1", Label: "Claude Opus 4.1 (Bedrock, deprecated 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, CacheReadPer1k: 0.0015, CacheCreationPer1k: 0.01875, ContextWindow: 200000},
+			{ID: "anthropic.claude-sonnet-4-6", Label: "Claude Sonnet 4.6 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
+			{ID: "anthropic.claude-sonnet-4-5", Label: "Claude Sonnet 4.5 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 200000},
+			{ID: "anthropic.claude-haiku-4-5", Label: "Claude Haiku 4.5 (Bedrock)", InputPer1k: 0.001, OutputPer1k: 0.005, CacheReadPer1k: 0.0001, CacheCreationPer1k: 0.00125, ContextWindow: 200000},
 			{ID: "meta.llama3-3-70b-instruct", Label: "Llama 3.3 70B (Bedrock)", InputPer1k: 0.00072, OutputPer1k: 0.00072, ContextWindow: 128000},
 			{ID: "amazon.nova-2-lite", Label: "Amazon Nova 2 Lite (Bedrock, preview)", InputPer1k: 0.0003, OutputPer1k: 0.0025, ContextWindow: 1000000},
 			{ID: "amazon.nova-pro", Label: "Amazon Nova Pro (Bedrock)", InputPer1k: 0.0008, OutputPer1k: 0.0032, ContextWindow: 300000},
@@ -358,6 +391,10 @@ var providers = []Provider{
 		AuthHeaderTemplate: "Bearer ${API_KEY}",
 		DefaultContentType: "application/json",
 		BrandColor:         "#4285F4",
+		// ParserID stays empty (path-style dispatch via IsVertexPathStyle);
+		// Anthropic-on-Vertex requests are metered under the "anthropic"
+		// surface with the bare, unversioned model id.
+		PricingSurfaces: []string{"anthropic"},
 		// Vertex carries the model in the URL path and authenticates with a
 		// service-account-minted OAuth token (api_key = "keyfile::<base64 SA>").
 		// Only Anthropic-on-Vertex is metered today: the request parser maps the
@@ -369,14 +406,14 @@ var providers = []Provider{
 		// exists — the router denies unmeterable publishers rather than forward
 		// them uncounted.
 		Models: []Model{
-			{ID: "claude-fable-5", Label: "Claude Fable 5 (Vertex)", InputPer1k: 0.010, OutputPer1k: 0.050, ContextWindow: 1000000},
-			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "claude-opus-4-6", Label: "Claude Opus 4.6 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
-			{ID: "claude-opus-4-1", Label: "Claude Opus 4.1 (Vertex, deprecated 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, ContextWindow: 200000},
-			{ID: "claude-sonnet-4-6", Label: "Claude Sonnet 4.6 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
-			{ID: "claude-sonnet-4-5", Label: "Claude Sonnet 4.5 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 200000},
-			{ID: "claude-haiku-4-5", Label: "Claude Haiku 4.5 (Vertex)", InputPer1k: 0.001, OutputPer1k: 0.005, ContextWindow: 200000},
+			{ID: "claude-fable-5", Label: "Claude Fable 5 (Vertex)", InputPer1k: 0.010, OutputPer1k: 0.050, CacheReadPer1k: 0.001, CacheCreationPer1k: 0.0125, ContextWindow: 1000000},
+			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "claude-opus-4-6", Label: "Claude Opus 4.6 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "claude-opus-4-1", Label: "Claude Opus 4.1 (Vertex, deprecated 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, CacheReadPer1k: 0.0015, CacheCreationPer1k: 0.01875, ContextWindow: 200000},
+			{ID: "claude-sonnet-4-6", Label: "Claude Sonnet 4.6 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
+			{ID: "claude-sonnet-4-5", Label: "Claude Sonnet 4.5 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 200000},
+			{ID: "claude-haiku-4-5", Label: "Claude Haiku 4.5 (Vertex)", InputPer1k: 0.001, OutputPer1k: 0.005, CacheReadPer1k: 0.0001, CacheCreationPer1k: 0.00125, ContextWindow: 200000},
 		},
 	},
 	{
@@ -390,6 +427,7 @@ var providers = []Provider{
 		DefaultContentType: "application/json",
 		BrandColor:         "#FF7000",
 		ParserID:           "openai",
+		PricingSurfaces:    []string{"openai"},
 		// Pricing + context windows cross-checked against LiteLLM. Key
 		// gotchas the marketing page hides:
 		//   - `mistral-medium-latest` aliases to Medium 3.1 ($0.40/$2),
@@ -448,6 +486,10 @@ var providers = []Provider{
 		// model id "k3") is account-bound seat licensing rather than a
 		// meterable platform key, so it's deliberately not the default.
 		ParserID: "",
+		// Both body shapes are metered: /v1/chat/completions under
+		// "openai", /anthropic/v1/messages under "anthropic" — so the
+		// K3 entry is priced on both surfaces.
+		PricingSurfaces: []string{"openai", "anthropic"},
 		// Pricing per Moonshot's platform rates at K3 launch (July 2026):
 		// $3/$15 per MTok with $0.30 cached input, flat across the 1M-token
 		// window. kimi-k3 is the ONLY model the platform serves newer
@@ -458,7 +500,14 @@ var providers = []Provider{
 		// The consumer app's "K3 Swarm Max" mode is not an API SKU, so it
 		// doesn't appear here.
 		Models: []Model{
-			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
+			// Carries both cache shapes: Moonshot reports cache hits
+			// OpenAI-style on /v1/chat/completions (CachedInputPer1k)
+			// and Anthropic-style on /anthropic/v1/messages
+			// (CacheReadPer1k) — $0.30/MTok either way. Each surface's
+			// cost formula reads only its own field, so the superset
+			// entry prices both endpoints correctly. No cache-creation
+			// rate published; writes bill at the input rate.
+			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, CachedInputPer1k: 0.0003, CacheReadPer1k: 0.0003, ContextWindow: 1000000},
 		},
 	},
 	{
@@ -758,13 +807,28 @@ func IsBedrockPathStyle(providerID string) bool {
 func (p Provider) ToAPIResponse() api.AgentNetworkCatalogProvider {
 	models := make([]api.AgentNetworkCatalogModel, 0, len(p.Models))
 	for _, m := range p.Models {
-		models = append(models, api.AgentNetworkCatalogModel{
+		am := api.AgentNetworkCatalogModel{
 			Id:            m.ID,
 			Label:         m.Label,
 			InputPer1k:    m.InputPer1k,
 			OutputPer1k:   m.OutputPer1k,
 			ContextWindow: m.ContextWindow,
-		})
+		}
+		// Cache rates are emitted only when configured so the dashboard
+		// can prefill them; 0 stays off the wire (absent = no rate).
+		if m.CachedInputPer1k > 0 {
+			v := m.CachedInputPer1k
+			am.CachedInputPer1k = &v
+		}
+		if m.CacheReadPer1k > 0 {
+			v := m.CacheReadPer1k
+			am.CacheReadPer1k = &v
+		}
+		if m.CacheCreationPer1k > 0 {
+			v := m.CacheCreationPer1k
+			am.CacheCreationPer1k = &v
+		}
+		models = append(models, am)
 	}
 	kind := api.AgentNetworkCatalogProviderKindProvider
 	switch p.Kind {
@@ -784,6 +848,10 @@ func (p Provider) ToAPIResponse() api.AgentNetworkCatalogProvider {
 		BrandColor:         p.BrandColor,
 		Models:             models,
 	}
+	if len(p.PricingSurfaces) > 0 {
+		surfaces := append([]string(nil), p.PricingSurfaces...)
+		resp.PricingSurfaces = &surfaces
+	}
 	if len(p.ExtraHeaders) > 0 {
 		extras := make([]api.AgentNetworkCatalogExtraHeader, 0, len(p.ExtraHeaders))
 		for _, h := range p.ExtraHeaders {

management/internals/modules/agentnetwork/handlers/providers_handler.go

@@ -7,6 +7,7 @@ package handlers
 
 import (
 	"encoding/json"
+	"math"
 	"net/http"
 	"net/url"
 	"strings"
@@ -15,6 +16,7 @@ import (
 
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	nbcontext "github.com/netbirdio/netbird/management/server/context"
 	"github.com/netbirdio/netbird/shared/management/http/api"
@@ -52,11 +54,45 @@ func (h *handler) getCatalogProviders(w http.ResponseWriter, r *http.Request) {
 	entries := catalog.All()
 	out := make([]api.AgentNetworkCatalogProvider, 0, len(entries))
 	for _, e := range entries {
-		out = append(out, e.ToAPIResponse())
+		resp := e.ToAPIResponse()
+		applyDefaultPricing(e, &resp)
+		out = append(out, resp)
 	}
 	util.WriteJSONObject(r.Context(), w, out)
 }
 
+// applyDefaultPricing overwrites the catalog response's model rates with
+// the LIVE default pricing table, which may differ from the compiled-in
+// catalog rates when the operator provides a defaults_llm_pricing.yaml.
+// This keeps the dashboard's model-row prefill identical to what the
+// proxy will actually bill — the same table the synthesizer ships.
+func applyDefaultPricing(cp catalog.Provider, resp *api.AgentNetworkCatalogProvider) {
+	if len(cp.PricingSurfaces) == 0 {
+		return
+	}
+	for i := range resp.Models {
+		m := &resp.Models[i]
+		e, ok := pricing.LookupDefault(cp.PricingSurfaces, m.Id)
+		if !ok {
+			continue
+		}
+		m.InputPer1k = e.InputPer1k
+		m.OutputPer1k = e.OutputPer1k
+		m.CachedInputPer1k = positiveRatePtr(e.CachedInputPer1k)
+		m.CacheReadPer1k = positiveRatePtr(e.CacheReadPer1k)
+		m.CacheCreationPer1k = positiveRatePtr(e.CacheCreationPer1k)
+	}
+}
+
+// positiveRatePtr renders a cache rate for the API: absent (nil) when
+// unset, matching the catalog response convention.
+func positiveRatePtr(v float64) *float64 {
+	if v <= 0 {
+		return nil
+	}
+	return &v
+}
+
 func (h *handler) getAllProviders(w http.ResponseWriter, r *http.Request) {
 	userAuth, err := nbcontext.GetUserAuthFromContext(r.Context())
 	if err != nil {
@@ -213,5 +249,38 @@ func validate(req *api.AgentNetworkProviderRequest, requireAPIKey bool) error {
 	if requireAPIKey && (req.ApiKey == nil || strings.TrimSpace(*req.ApiKey) == "") {
 		return status.Errorf(status.InvalidArgument, "api_key is required")
 	}
+	if req.Models != nil {
+		for i, m := range *req.Models {
+			if err := validateModel(i, m); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+// validateModel is the single ingress guard for operator-entered pricing:
+// these rates are synthesized into the proxy's cost_meter config verbatim,
+// and a negative or non-finite rate there would poison every cost the
+// proxy records, so reject at the API boundary.
+func validateModel(i int, m api.AgentNetworkProviderModel) error {
+	if strings.TrimSpace(m.Id) == "" {
+		return status.Errorf(status.InvalidArgument, "models[%d]: id is required", i)
+	}
+	rates := map[string]*float64{
+		"input_per_1k":          &m.InputPer1k,
+		"output_per_1k":         &m.OutputPer1k,
+		"cached_input_per_1k":   m.CachedInputPer1k,
+		"cache_read_per_1k":     m.CacheReadPer1k,
+		"cache_creation_per_1k": m.CacheCreationPer1k,
+	}
+	for field, v := range rates {
+		if v == nil {
+			continue
+		}
+		if *v < 0 || math.IsNaN(*v) || math.IsInf(*v, 0) {
+			return status.Errorf(status.InvalidArgument, "models[%d] (%s): %s must be a finite, non-negative USD rate", i, m.Id, field)
+		}
+	}
 	return nil
 }

management/internals/modules/agentnetwork/handlers/providers_handler_test.go

@@ -0,0 +1,53 @@
+package handlers
+
+import (
+	"math"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+func f(v float64) *float64 { return &v }
+
+// TestValidate_ModelRates guards the single ingress point for operator-entered
+// pricing. These rates flow verbatim into the proxy's cost_meter config at
+// synthesis time; the proxy treats a bad rate as a chain-build failure, so
+// rejecting here is what keeps an account's gateway from going down.
+func TestValidate_ModelRates(t *testing.T) {
+	base := func(models ...api.AgentNetworkProviderModel) *api.AgentNetworkProviderRequest {
+		key := "sk-test"
+		return &api.AgentNetworkProviderRequest{
+			ProviderId:  "openai_api",
+			Name:        "OpenAI",
+			UpstreamUrl: "https://api.openai.com",
+			ApiKey:      &key,
+			Models:      &models,
+		}
+	}
+
+	valid := api.AgentNetworkProviderModel{
+		Id: "gpt-4o", InputPer1k: 0.0025, OutputPer1k: 0.01,
+		CachedInputPer1k: f(0.00125),
+	}
+	require.NoError(t, validate(base(valid), true), "finite non-negative rates must pass")
+
+	zeroRates := api.AgentNetworkProviderModel{Id: "self-hosted-llama", InputPer1k: 0, OutputPer1k: 0}
+	require.NoError(t, validate(base(zeroRates), true), "explicit zero prices are allowed (free / self-hosted models)")
+
+	cases := map[string]api.AgentNetworkProviderModel{
+		"empty id":           {Id: "  ", InputPer1k: 0.001, OutputPer1k: 0.002},
+		"negative input":     {Id: "m", InputPer1k: -0.001, OutputPer1k: 0.002},
+		"negative output":    {Id: "m", InputPer1k: 0.001, OutputPer1k: -0.002},
+		"NaN input":          {Id: "m", InputPer1k: math.NaN(), OutputPer1k: 0.002},
+		"Inf output":         {Id: "m", InputPer1k: 0.001, OutputPer1k: math.Inf(1)},
+		"negative cached":    {Id: "m", InputPer1k: 0.001, OutputPer1k: 0.002, CachedInputPer1k: f(-1)},
+		"NaN cache read":     {Id: "m", InputPer1k: 0.001, OutputPer1k: 0.002, CacheReadPer1k: f(math.NaN())},
+		"Inf cache creation": {Id: "m", InputPer1k: 0.001, OutputPer1k: 0.002, CacheCreationPer1k: f(math.Inf(-1))},
+	}
+	for name, m := range cases {
+		assert.Error(t, validate(base(m), true), "case %q must be rejected", name)
+	}
+}

management/internals/modules/agentnetwork/pricing/defaults.go

@@ -0,0 +1,156 @@
+// Package pricing builds the default LLM pricing table the synthesizer
+// ships to the proxy's cost_meter middleware. The catalog is the single
+// source of default rates: every catalog provider's models are folded
+// into the pricing surfaces the provider declares (PricingSurfaces),
+// then a small supplemental list adds priced-but-not-operator-selectable
+// entries. Management is the sole pricing authority — the proxy carries
+// no embedded price list and bills exclusively from the table it is
+// sent.
+//go:generate go run gen.go
+
+package pricing
+
+import (
+	"sync"
+	"sync/atomic"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
+)
+
+// Entry is a single model's pricing in USD per 1k tokens. This struct IS
+// the wire shape: the synthesizer marshals it verbatim into cost_meter's
+// ConfigJSON, and the proxy unmarshals the same field names.
+//
+// A zero rate means "no rate configured" — the proxy bills that cache
+// bucket at InputPer1k (identical semantics to the retired proxy-embedded
+// table). CachedInputPer1k is the OpenAI shape (cached prompt tokens are
+// a subset of input); CacheReadPer1k / CacheCreationPer1k are the
+// Anthropic shape (additive buckets).
+type Entry struct {
+	InputPer1k         float64 `json:"input_per_1k"`
+	OutputPer1k        float64 `json:"output_per_1k"`
+	CachedInputPer1k   float64 `json:"cached_input_per_1k,omitempty"`
+	CacheReadPer1k     float64 `json:"cache_read_per_1k,omitempty"`
+	CacheCreationPer1k float64 `json:"cache_creation_per_1k,omitempty"`
+}
+
+// supplementalDefaults are (surface, model) entries that are priced but
+// deliberately not operator-selectable in the catalog. Each carries a
+// reason; when one of these models joins a catalog lineup, delete the
+// row here — the collision test fails loudly if the rates ever disagree.
+var supplementalDefaults = map[string]map[string]Entry{
+	"openai": {
+		// GPT-5 (2025) family — kept for gateway requests using the
+		// unsuffixed ids; the dashboard offers only the 5.x lineup.
+		"gpt-5":      {InputPer1k: 0.00125, OutputPer1k: 0.01, CachedInputPer1k: 0.000125},
+		"gpt-5-mini": {InputPer1k: 0.00025, OutputPer1k: 0.002, CachedInputPer1k: 0.000025},
+		"gpt-5-nano": {InputPer1k: 0.00005, OutputPer1k: 0.0004, CachedInputPer1k: 0.000005},
+	},
+	"anthropic": {
+		// claude-opus-5 is not yet in the catalog lineup but gateway /
+		// grandfathered traffic uses it; priced so it isn't skipped.
+		"claude-opus-5": {InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625},
+		// "kimi-k3[1m]" is the 1M-context alias some Claude Code guides
+		// configure against Moonshot's Anthropic-compatible endpoint;
+		// priced identically to kimi-k3 so those requests aren't skipped.
+		"kimi-k3[1m]": {InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003},
+	},
+	"bedrock": {
+		"anthropic.claude-opus-5": {InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625},
+	},
+}
+
+var (
+	compiledOnce  sync.Once
+	compiledTable map[string]map[string]Entry
+	// mergedTable holds the current live table when a pricing defaults
+	// file is loaded: the file merged entry-whole over the compiled-in
+	// base. Nil while no file is loaded (or after the file is removed),
+	// in which case the compiled-in table serves. Swapped atomically by
+	// the file loader/reloader; readers never block.
+	mergedTable atomic.Pointer[map[string]map[string]Entry]
+)
+
+// DefaultTable returns the current default pricing table keyed
+// surface -> model -> Entry: the management-side defaults file (see
+// LoadFile / StartReloader) when one is loaded, merged over the
+// compiled-in catalog table, which alone serves as the fallback when no
+// file exists. The snapshot may change between calls as the file is
+// re-read — consumers (the synthesizer on every reconcile, the catalog
+// endpoint on every request) pick up fresh rates automatically. Callers
+// must not mutate the returned maps.
+func DefaultTable() map[string]map[string]Entry {
+	if t := mergedTable.Load(); t != nil {
+		return *t
+	}
+	return compiledBase()
+}
+
+// compiledBase returns the compiled-in table (catalog + supplementals),
+// built once.
+func compiledBase() map[string]map[string]Entry {
+	compiledOnce.Do(func() {
+		compiledTable = buildDefaultTable()
+	})
+	return compiledTable
+}
+
+func buildDefaultTable() map[string]map[string]Entry {
+	out := make(map[string]map[string]Entry)
+	for _, p := range catalog.All() {
+		for _, surface := range p.PricingSurfaces {
+			inner, ok := out[surface]
+			if !ok {
+				inner = make(map[string]Entry)
+				out[surface] = inner
+			}
+			for _, m := range p.Models {
+				// First writer wins; providers contributing the same
+				// (surface, model) must agree on rates — enforced by
+				// TestDefaultTable_NoConflictingContributions.
+				if _, dup := inner[m.ID]; dup {
+					continue
+				}
+				inner[m.ID] = entryFromCatalogModel(m)
+			}
+		}
+	}
+	for surface, models := range supplementalDefaults {
+		inner, ok := out[surface]
+		if !ok {
+			inner = make(map[string]Entry)
+			out[surface] = inner
+		}
+		for id, e := range models {
+			if _, dup := inner[id]; dup {
+				continue
+			}
+			inner[id] = e
+		}
+	}
+	return out
+}
+
+func entryFromCatalogModel(m catalog.Model) Entry {
+	return Entry{
+		InputPer1k:         m.InputPer1k,
+		OutputPer1k:        m.OutputPer1k,
+		CachedInputPer1k:   m.CachedInputPer1k,
+		CacheReadPer1k:     m.CacheReadPer1k,
+		CacheCreationPer1k: m.CacheCreationPer1k,
+	}
+}
+
+// LookupDefault returns the default entry for model on the first of the
+// given surfaces that prices it. Used by the synthesizer to seed a
+// per-provider entry with default cache rates before overlaying the
+// operator's stored prices.
+func LookupDefault(surfaces []string, model string) (Entry, bool) {
+	table := DefaultTable()
+	for _, s := range surfaces {
+		if e, ok := table[s][model]; ok {
+			return e, true
+		}
+	}
+	return Entry{}, false
+}

management/internals/modules/agentnetwork/pricing/defaults_llm_pricing.example.yaml

@@ -1,83 +1,176 @@
-# Embedded default pricing for llm_observability. Compiled into the proxy
-# binary via go:embed in pricing.go; cost annotation works out of the box
-# without any operator action.
+# Default LLM pricing used by NetBird's Agent Network cost metering.
+# GENERATED from the management catalog — do not edit this file in the
+# repository; regenerate with:
 #
-# Operators override entries by dropping a pricing.yaml into --plugin-data-dir
-# (or whichever basename is given via params.pricing_path). The override file
-# only needs entries the operator wants to change; missing entries fall
-# through to these defaults.
+#   go generate ./management/internals/modules/agentnetwork/pricing
 #
-# Values are USD per 1_000 tokens. Public list prices drift; ship a fresh
-# binary or override individual entries via the override file as needed.
+# Operators: copy this file to <datadir>/defaults_llm_pricing.yaml (or
+# any path configured via management.json:
 #
-# Optional cache fields:
-#   cached_input_per_1k    OpenAI: rate for prompt_tokens_details.cached_tokens
-#                          (a SUBSET of prompt_tokens). Typically 0.5x input.
-#                          Absent → cached portion bills at input_per_1k.
-#   cache_read_per_1k      Anthropic: rate for cache_read_input_tokens
-#                          (ADDITIVE to input_tokens). Typically 0.1x input.
-#                          Absent → cache reads bill at input_per_1k.
-#   cache_creation_per_1k  Anthropic: rate for cache_creation_input_tokens
-#                          (ADDITIVE to input_tokens). Typically 1.25x input.
-#                          Absent → cache writes bill at input_per_1k.
+#   { "AgentNetwork": { "PricingDefaultsFile": "/path/defaults_llm_pricing.yaml" } }
+#
+# ) and adjust the entries you want to change. Management re-reads the
+# file periodically (mtime poll, every minute): the live table feeds the
+# proxies' cost metering and the dashboard's model-price prefill, so
+# edits apply without a restart. Your file only needs the entries you
+# want to change — but each entry REPLACES the built-in entry for that
+# surface+model whole, so repeat the cache rates you want to keep.
+# Unknown fields and negative or non-finite rates are rejected: at
+# startup that fails boot (for an explicitly configured path); at
+# runtime the previous table is kept and a warning is logged. Deleting
+# the file reverts to the built-in defaults below.
+#
+# Top-level keys are pricing surfaces — the parser shape requests are
+# metered under: "openai" (also Azure, Mistral, and OpenAI-compatible
+# gateways), "anthropic" (also Anthropic-on-Vertex), "bedrock"
+# (normalized ids, e.g. anthropic.claude-sonnet-4-5). Model keys must be
+# the normalized id the proxy meters (version/region suffixes stripped).
+#
+# Values are USD per 1_000 tokens. Optional cache fields:
+#   cached_input_per_1k    OpenAI shape: rate for cached prompt tokens
+#                          (a SUBSET of input tokens). Absent -> cached
+#                          portion bills at input_per_1k.
+#   cache_read_per_1k      Anthropic shape: rate for cache_read tokens
+#                          (ADDITIVE to input). Absent -> input rate.
+#   cache_creation_per_1k  Anthropic shape: rate for cache_creation
+#                          tokens (ADDITIVE to input). Absent -> input
+#                          rate.
+
+anthropic:
+  claude-fable-5:
+    input_per_1k: 0.01
+    output_per_1k: 0.05
+    cache_read_per_1k: 0.001
+    cache_creation_per_1k: 0.0125
+  claude-haiku-4-5:
+    input_per_1k: 0.001
+    output_per_1k: 0.005
+    cache_read_per_1k: 0.0001
+    cache_creation_per_1k: 0.00125
+  claude-opus-4-1:
+    input_per_1k: 0.015
+    output_per_1k: 0.075
+    cache_read_per_1k: 0.0015
+    cache_creation_per_1k: 0.01875
+  claude-opus-4-6:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  claude-opus-4-7:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  claude-opus-4-8:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  claude-opus-5:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  claude-sonnet-4-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+  claude-sonnet-4-6:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cached_input_per_1k: 0.0003
+    cache_read_per_1k: 0.0003
+  "kimi-k3[1m]":
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+
+bedrock:
+  amazon.nova-2-lite:
+    input_per_1k: 0.0003
+    output_per_1k: 0.0025
+  amazon.nova-lite:
+    input_per_1k: 0.00006
+    output_per_1k: 0.00024
+  amazon.nova-micro:
+    input_per_1k: 0.000035
+    output_per_1k: 0.00014
+  amazon.nova-pro:
+    input_per_1k: 0.0008
+    output_per_1k: 0.0032
+  anthropic.claude-haiku-4-5:
+    input_per_1k: 0.001
+    output_per_1k: 0.005
+    cache_read_per_1k: 0.0001
+    cache_creation_per_1k: 0.00125
+  anthropic.claude-opus-4-1:
+    input_per_1k: 0.015
+    output_per_1k: 0.075
+    cache_read_per_1k: 0.0015
+    cache_creation_per_1k: 0.01875
+  anthropic.claude-opus-4-6:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  anthropic.claude-opus-4-7:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  anthropic.claude-opus-4-8:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  anthropic.claude-opus-5:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  anthropic.claude-sonnet-4-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+  anthropic.claude-sonnet-4-6:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+  meta.llama3-3-70b-instruct:
+    input_per_1k: 0.00072
+    output_per_1k: 0.00072
 
 openai:
-  # OpenAI + OpenAI-compatible providers (openai_api, azure_openai_api,
-  # mistral_api, and the openai-parser gateways) all emit llm.provider="openai",
-  # so their models are priced here. Kept in sync with the management catalog;
-  # rates cross-checked against LiteLLM model_prices_and_context_window.json.
-
-  # GPT-5.x family — cache reads 10% of input (0.1x).
-  gpt-5.5:
-    input_per_1k: 0.005
-    output_per_1k: 0.03
-    cached_input_per_1k: 0.0005
-  gpt-5.5-pro:
-    input_per_1k: 0.03
-    output_per_1k: 0.18
-    cached_input_per_1k: 0.003
-  gpt-5.4:
-    input_per_1k: 0.0025
-    output_per_1k: 0.015
-    cached_input_per_1k: 0.00025
-  gpt-5.4-pro:
-    input_per_1k: 0.03
-    output_per_1k: 0.18
-    cached_input_per_1k: 0.003
-  gpt-5.4-mini:
-    input_per_1k: 0.00075
-    output_per_1k: 0.0045
-    cached_input_per_1k: 0.000075
-  gpt-5.4-nano:
-    input_per_1k: 0.0002
-    output_per_1k: 0.00125
-    cached_input_per_1k: 0.00002
-  gpt-5.3-codex:
-    input_per_1k: 0.00175
-    output_per_1k: 0.014
-    cached_input_per_1k: 0.000175
-  gpt-5.3-chat-latest:
-    input_per_1k: 0.00175
-    output_per_1k: 0.014
-    cached_input_per_1k: 0.000175
-  # GPT-5 (2025) family — kept for gateway requests using the unsuffixed ids.
-  gpt-5:
-    input_per_1k: 0.00125
-    output_per_1k: 0.01
-    cached_input_per_1k: 0.000125
-  gpt-5-mini:
-    input_per_1k: 0.00025
+  codestral-2508:
+    input_per_1k: 0.0003
+    output_per_1k: 0.0009
+  codestral-latest:
+    input_per_1k: 0.001
+    output_per_1k: 0.003
+  devstral-medium-latest:
+    input_per_1k: 0.0004
     output_per_1k: 0.002
-    cached_input_per_1k: 0.000025
-  gpt-5-nano:
-    input_per_1k: 0.00005
-    output_per_1k: 0.0004
-    cached_input_per_1k: 0.000005
-  o4-mini:
-    input_per_1k: 0.0011
-    output_per_1k: 0.0044
-    cached_input_per_1k: 0.000275
-  # GPT-4.1 family — cache reads 25% of input.
+  devstral-small-latest:
+    input_per_1k: 0.0001
+    output_per_1k: 0.0003
+  gpt-3.5-turbo:
+    input_per_1k: 0.0005
+    output_per_1k: 0.0015
+  gpt-35-turbo:
+    input_per_1k: 0.0005
+    output_per_1k: 0.0015
+  gpt-4-turbo:
+    input_per_1k: 0.01
+    output_per_1k: 0.03
   gpt-4.1:
     input_per_1k: 0.002
     output_per_1k: 0.008
@@ -90,7 +183,6 @@ openai:
     input_per_1k: 0.0001
     output_per_1k: 0.0004
     cached_input_per_1k: 0.000025
-  # GPT-4o family — cache reads 50% of input (0.5x).
   gpt-4o:
     input_per_1k: 0.0025
     output_per_1k: 0.01
@@ -99,200 +191,92 @@ openai:
     input_per_1k: 0.00015
     output_per_1k: 0.0006
     cached_input_per_1k: 0.000075
-  # Older GPT — no prompt caching.
-  gpt-4-turbo:
-    input_per_1k: 0.01
-    output_per_1k: 0.03
-  gpt-3.5-turbo:
-    input_per_1k: 0.0005
-    output_per_1k: 0.0015
-  gpt-35-turbo: # Azure deployment alias of gpt-3.5-turbo
-    input_per_1k: 0.0005
-    output_per_1k: 0.0015
-  # Embeddings — no caching, no output tokens.
-  text-embedding-3-large:
-    input_per_1k: 0.00013
-    output_per_1k: 0
-  text-embedding-3-small:
-    input_per_1k: 0.00002
-    output_per_1k: 0
-
-  # Mistral (mistral_api) — routed via the openai parser; no prompt caching.
-  mistral-large-latest:
-    input_per_1k: 0.0005
-    output_per_1k: 0.0015
-  mistral-medium-latest:
-    input_per_1k: 0.0004
+  gpt-5:
+    input_per_1k: 0.00125
+    output_per_1k: 0.01
+    cached_input_per_1k: 0.000125
+  gpt-5-mini:
+    input_per_1k: 0.00025
     output_per_1k: 0.002
-  mistral-medium-3-5:
-    input_per_1k: 0.0015
-    output_per_1k: 0.0075
-  mistral-small-latest:
-    input_per_1k: 0.00006
-    output_per_1k: 0.00018
+    cached_input_per_1k: 0.000025
+  gpt-5-nano:
+    input_per_1k: 0.00005
+    output_per_1k: 0.0004
+    cached_input_per_1k: 0.000005
+  gpt-5.3-chat-latest:
+    input_per_1k: 0.00175
+    output_per_1k: 0.014
+    cached_input_per_1k: 0.000175
+  gpt-5.3-codex:
+    input_per_1k: 0.00175
+    output_per_1k: 0.014
+    cached_input_per_1k: 0.000175
+  gpt-5.4:
+    input_per_1k: 0.0025
+    output_per_1k: 0.015
+    cached_input_per_1k: 0.00025
+  gpt-5.4-mini:
+    input_per_1k: 0.00075
+    output_per_1k: 0.0045
+    cached_input_per_1k: 0.000075
+  gpt-5.4-nano:
+    input_per_1k: 0.0002
+    output_per_1k: 0.00125
+    cached_input_per_1k: 0.00002
+  gpt-5.4-pro:
+    input_per_1k: 0.03
+    output_per_1k: 0.18
+    cached_input_per_1k: 0.003
+  gpt-5.5:
+    input_per_1k: 0.005
+    output_per_1k: 0.03
+    cached_input_per_1k: 0.0005
+  gpt-5.5-pro:
+    input_per_1k: 0.03
+    output_per_1k: 0.18
+    cached_input_per_1k: 0.003
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cached_input_per_1k: 0.0003
+    cache_read_per_1k: 0.0003
   magistral-medium-latest:
     input_per_1k: 0.002
     output_per_1k: 0.005
   magistral-small-latest:
     input_per_1k: 0.0005
     output_per_1k: 0.0015
-  devstral-medium-latest:
-    input_per_1k: 0.0004
-    output_per_1k: 0.002
-  devstral-small-latest:
-    input_per_1k: 0.0001
-    output_per_1k: 0.0003
-  codestral-2508:
-    input_per_1k: 0.0003
-    output_per_1k: 0.0009
-  codestral-latest:
-    input_per_1k: 0.001
-    output_per_1k: 0.003
   ministral-3-14b-2512:
     input_per_1k: 0.0002
     output_per_1k: 0.0002
-  ministral-8b-latest:
-    input_per_1k: 0.00015
-    output_per_1k: 0.00015
   ministral-3-3b-2512:
     input_per_1k: 0.0001
     output_per_1k: 0.0001
+  ministral-8b-latest:
+    input_per_1k: 0.00015
+    output_per_1k: 0.00015
   mistral-embed:
     input_per_1k: 0.0001
     output_per_1k: 0
-
-  # Kimi / Moonshot AI (kimi_api) — OpenAI-compatible /v1 endpoint. Moonshot
-  # reports cache hits OpenAI-style when present; cached input is 10% of
-  # input ($0.30 vs $3.00 per MTok). kimi-k3 is the only model the platform
-  # serves newer accounts (K2-era ids and kimi-latest 404), matching the
-  # management catalog.
-  kimi-k3:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cached_input_per_1k: 0.0003
-
-anthropic:
-  # Claude 4.x family — cache reads ≈10% of input, cache writes ≈125% of input.
-  # Pricing source: Anthropic's current published rates per million tokens,
-  # divided by 1000 for the per-1k figures stored here.
-  claude-fable-5:
-    input_per_1k: 0.010
-    output_per_1k: 0.050
-    cache_read_per_1k: 0.001
-    cache_creation_per_1k: 0.0125
-  claude-opus-5:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  claude-opus-4-8:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  claude-opus-4-7:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  claude-opus-4-6:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  claude-opus-4-1:
-    input_per_1k: 0.015
-    output_per_1k: 0.075
-    cache_read_per_1k: 0.0015
-    cache_creation_per_1k: 0.01875
-  claude-sonnet-4-6:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-  claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-  claude-haiku-4-5:
-    input_per_1k: 0.001
-    output_per_1k: 0.005
-    cache_read_per_1k: 0.0001
-    cache_creation_per_1k: 0.00125
-
-  # Kimi / Moonshot AI (kimi_api) via the Anthropic-compatible endpoint
-  # (/anthropic/v1/messages — the official Claude Code setup). Same rates
-  # as the OpenAI-shape entry above. "kimi-k3[1m]" is the model id some
-  # Claude Code guides set for the 1M-context alias; priced identically so
-  # cost metering doesn't silently skip those requests.
-  kimi-k3:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-  "kimi-k3[1m]":
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-
-bedrock:
-  # AWS Bedrock model ids, normalised by the request parser (cross-region
-  # inference-profile prefix + version/throughput suffix stripped), e.g.
-  # eu.anthropic.claude-sonnet-4-5-20250929-v1:0 -> anthropic.claude-sonnet-4-5.
-  # Anthropic-on-Bedrock keeps the additive cache buckets (read ≈0.1x input,
-  # write ≈1.25x input); Nova / Llama report no cache, so cost is input+output.
-  anthropic.claude-opus-5:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  anthropic.claude-opus-4-8:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  anthropic.claude-opus-4-7:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  anthropic.claude-opus-4-6:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  anthropic.claude-opus-4-1:
-    input_per_1k: 0.015
-    output_per_1k: 0.075
-    cache_read_per_1k: 0.0015
-    cache_creation_per_1k: 0.01875
-  anthropic.claude-sonnet-4-6:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-  anthropic.claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-  anthropic.claude-haiku-4-5:
-    input_per_1k: 0.001
-    output_per_1k: 0.005
-    cache_read_per_1k: 0.0001
-    cache_creation_per_1k: 0.00125
-  meta.llama3-3-70b-instruct:
-    input_per_1k: 0.00072
-    output_per_1k: 0.00072
-  amazon.nova-2-lite:
-    input_per_1k: 0.0003
-    output_per_1k: 0.0025
-  amazon.nova-pro:
-    input_per_1k: 0.0008
-    output_per_1k: 0.0032
-  amazon.nova-lite:
+  mistral-large-latest:
+    input_per_1k: 0.0005
+    output_per_1k: 0.0015
+  mistral-medium-3-5:
+    input_per_1k: 0.0015
+    output_per_1k: 0.0075
+  mistral-medium-latest:
+    input_per_1k: 0.0004
+    output_per_1k: 0.002
+  mistral-small-latest:
     input_per_1k: 0.00006
-    output_per_1k: 0.00024
-  amazon.nova-micro:
-    input_per_1k: 0.000035
-    output_per_1k: 0.00014
+    output_per_1k: 0.00018
+  o4-mini:
+    input_per_1k: 0.0011
+    output_per_1k: 0.0044
+    cached_input_per_1k: 0.000275
+  text-embedding-3-large:
+    input_per_1k: 0.00013
+    output_per_1k: 0
+  text-embedding-3-small:
+    input_per_1k: 0.00002
+    output_per_1k: 0

management/internals/modules/agentnetwork/pricing/defaults_test.go

@@ -0,0 +1,148 @@
+package pricing
+
+import (
+	"math"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
+)
+
+// TestDefaultTable_CoversEveryCatalogModel replaces the proxy's old
+// hand-maintained coverage list: because the table is built FROM the
+// catalog, drift is impossible by construction — this test guards the
+// fold itself (every catalog model of every surfaced provider resolves,
+// with exactly the catalog's rates).
+func TestDefaultTable_CoversEveryCatalogModel(t *testing.T) {
+	table := DefaultTable()
+	for _, p := range catalog.All() {
+		if len(p.PricingSurfaces) == 0 {
+			assert.Empty(t, p.Models, "catalog entry %s declares models but no pricing surfaces — those models would never be priced", p.ID)
+			continue
+		}
+		for _, surface := range p.PricingSurfaces {
+			byModel, ok := table[surface]
+			require.True(t, ok, "surface %q (provider %s) missing from default table", surface, p.ID)
+			for _, m := range p.Models {
+				e, ok := byModel[m.ID]
+				require.True(t, ok, "%s/%s (provider %s) missing from default table", surface, m.ID, p.ID)
+				assert.Equal(t, m.InputPer1k, e.InputPer1k, "%s/%s input rate", surface, m.ID)
+				assert.Equal(t, m.OutputPer1k, e.OutputPer1k, "%s/%s output rate", surface, m.ID)
+			}
+		}
+	}
+}
+
+// TestDefaultTable_NoConflictingContributions enforces the collision rule
+// documented on catalog.Provider.PricingSurfaces: when two catalog
+// providers contribute the same (surface, model) pair — azure/vertex
+// mirroring openai/anthropic, kimi on both surfaces — their rates must be
+// identical, because the surface-keyed table can only hold one entry.
+// If a provider ever diverges (e.g. Azure reprices a model), this fails
+// and the divergence must move to per-provider-record pricing.
+func TestDefaultTable_NoConflictingContributions(t *testing.T) {
+	type contribution struct {
+		providerID string
+		entry      Entry
+	}
+	seen := map[string]map[string]contribution{}
+	for _, p := range catalog.All() {
+		for _, surface := range p.PricingSurfaces {
+			if seen[surface] == nil {
+				seen[surface] = map[string]contribution{}
+			}
+			for _, m := range p.Models {
+				e := entryFromCatalogModel(m)
+				if prev, dup := seen[surface][m.ID]; dup {
+					assert.Equal(t, prev.entry, e,
+						"%s/%s: %s and %s contribute different rates", surface, m.ID, prev.providerID, p.ID)
+					continue
+				}
+				seen[surface][m.ID] = contribution{providerID: p.ID, entry: e}
+			}
+		}
+	}
+	// Supplemental entries must never shadow a catalog-contributed model —
+	// they exist precisely because the catalog does NOT list them.
+	for surface, models := range supplementalDefaults {
+		for id := range models {
+			_, fromCatalog := seen[surface][id]
+			assert.False(t, fromCatalog, "supplemental %s/%s is now in the catalog — delete the supplemental row", surface, id)
+		}
+	}
+}
+
+// TestDefaultTable_AllRatesFiniteNonNegative mirrors the proxy-side
+// NewTable validation so a bad catalog edit is caught here, at unit-test
+// time, rather than as a chain-build failure on every proxy.
+func TestDefaultTable_AllRatesFiniteNonNegative(t *testing.T) {
+	for surface, models := range DefaultTable() {
+		for id, e := range models {
+			for field, v := range map[string]float64{
+				"input":          e.InputPer1k,
+				"output":         e.OutputPer1k,
+				"cached_input":   e.CachedInputPer1k,
+				"cache_read":     e.CacheReadPer1k,
+				"cache_creation": e.CacheCreationPer1k,
+			} {
+				assert.False(t, v < 0 || math.IsNaN(v) || math.IsInf(v, 0),
+					"%s/%s: %s rate %v must be finite and non-negative", surface, id, field, v)
+			}
+		}
+	}
+}
+
+// TestDefaultTable_PinnedRates pins rates that previously drifted or are
+// easy to mis-enter (carried over from the proxy's retired
+// defaults_coverage_test), plus the supplemental entries.
+func TestDefaultTable_PinnedRates(t *testing.T) {
+	table := DefaultTable()
+
+	gpt54 := table["openai"]["gpt-5.4"]
+	assert.InDelta(t, 0.0025, gpt54.InputPer1k, 1e-9, "gpt-5.4 input")
+	assert.InDelta(t, 0.015, gpt54.OutputPer1k, 1e-9, "gpt-5.4 output")
+	assert.InDelta(t, 0.00025, gpt54.CachedInputPer1k, 1e-9, "gpt-5.4 cached input")
+
+	sonnet := table["bedrock"]["anthropic.claude-sonnet-4-5"]
+	assert.InDelta(t, 0.003, sonnet.InputPer1k, 1e-9, "bedrock sonnet-4-5 input")
+	assert.InDelta(t, 0.015, sonnet.OutputPer1k, 1e-9, "bedrock sonnet-4-5 output")
+	assert.InDelta(t, 0.0003, sonnet.CacheReadPer1k, 1e-9, "bedrock sonnet-4-5 cache read")
+	assert.InDelta(t, 0.00375, sonnet.CacheCreationPer1k, 1e-9, "bedrock sonnet-4-5 cache creation")
+
+	// Vertex Claude prices under "anthropic" with the bare id.
+	fable := table["anthropic"]["claude-fable-5"]
+	assert.InDelta(t, 0.010, fable.InputPer1k, 1e-9, "claude-fable-5 input")
+	assert.InDelta(t, 0.0125, fable.CacheCreationPer1k, 1e-9, "claude-fable-5 cache creation")
+
+	// Supplementals present on their surfaces.
+	for surface, ids := range map[string][]string{
+		"openai":    {"gpt-5", "gpt-5-mini", "gpt-5-nano"},
+		"anthropic": {"claude-opus-5", "kimi-k3[1m]", "kimi-k3"},
+		"bedrock":   {"anthropic.claude-opus-5"},
+	} {
+		for _, id := range ids {
+			_, ok := table[surface][id]
+			assert.True(t, ok, "%s/%s must be priced", surface, id)
+		}
+	}
+
+	// Embeddings bill input-only — output stays zero.
+	emb := table["openai"]["text-embedding-3-large"]
+	assert.Zero(t, emb.OutputPer1k, "embedding output rate must be zero")
+	assert.Positive(t, emb.InputPer1k, "embedding input rate must be set")
+}
+
+func TestLookupDefault_SurfaceOrder(t *testing.T) {
+	// kimi-k3 exists on both surfaces; first surface in the slice wins.
+	e, ok := LookupDefault([]string{"openai", "anthropic"}, "kimi-k3")
+	require.True(t, ok)
+	assert.InDelta(t, 0.003, e.InputPer1k, 1e-9)
+
+	_, ok = LookupDefault([]string{"bedrock"}, "gpt-4o")
+	assert.False(t, ok, "gpt-4o is not a bedrock model")
+
+	_, ok = LookupDefault(nil, "gpt-4o")
+	assert.False(t, ok, "no surfaces, no match")
+}

management/internals/modules/agentnetwork/pricing/exampleyaml.go

@@ -0,0 +1,109 @@
+package pricing
+
+import (
+	"bytes"
+	"fmt"
+	"sort"
+	"strconv"
+)
+
+// MarshalDefaultsYAML renders the built-in default pricing table (catalog
+// + supplementals, WITHOUT any operator override) as the YAML schema
+// LoadOverrideFile consumes. It backs the generated
+// defaults_llm_pricing.example.yaml so operators start from a file that
+// matches the compiled-in rates exactly; a golden test keeps the two in
+// sync. Output is deterministic (sorted surfaces and models).
+func MarshalDefaultsYAML() []byte {
+	var b bytes.Buffer
+	b.WriteString(exampleHeader)
+
+	table := buildDefaultTable()
+	surfaces := make([]string, 0, len(table))
+	for s := range table {
+		surfaces = append(surfaces, s)
+	}
+	sort.Strings(surfaces)
+
+	for _, surface := range surfaces {
+		fmt.Fprintf(&b, "\n%s:\n", surface)
+		models := table[surface]
+		ids := make([]string, 0, len(models))
+		for id := range models {
+			ids = append(ids, id)
+		}
+		sort.Strings(ids)
+		for _, id := range ids {
+			e := models[id]
+			fmt.Fprintf(&b, "  %s:\n", yamlKey(id))
+			fmt.Fprintf(&b, "    input_per_1k: %s\n", rate(e.InputPer1k))
+			fmt.Fprintf(&b, "    output_per_1k: %s\n", rate(e.OutputPer1k))
+			if e.CachedInputPer1k > 0 {
+				fmt.Fprintf(&b, "    cached_input_per_1k: %s\n", rate(e.CachedInputPer1k))
+			}
+			if e.CacheReadPer1k > 0 {
+				fmt.Fprintf(&b, "    cache_read_per_1k: %s\n", rate(e.CacheReadPer1k))
+			}
+			if e.CacheCreationPer1k > 0 {
+				fmt.Fprintf(&b, "    cache_creation_per_1k: %s\n", rate(e.CacheCreationPer1k))
+			}
+		}
+	}
+	return b.Bytes()
+}
+
+// rate renders a USD-per-1k rate without float noise ("0.00015", not
+// "0.000150000000...").
+func rate(v float64) string {
+	return strconv.FormatFloat(v, 'f', -1, 64)
+}
+
+// yamlKey quotes model ids that YAML would otherwise misparse (e.g.
+// "kimi-k3[1m]" starts a flow sequence unquoted).
+func yamlKey(id string) string {
+	for _, r := range id {
+		switch r {
+		case '[', ']', '{', '}', ':', '#', ',', '&', '*', '!', '|', '>', '\'', '"', '%', '@', '`':
+			return strconv.Quote(id)
+		}
+	}
+	return id
+}
+
+const exampleHeader = `# Default LLM pricing used by NetBird's Agent Network cost metering.
+# GENERATED from the management catalog — do not edit this file in the
+# repository; regenerate with:
+#
+#   go generate ./management/internals/modules/agentnetwork/pricing
+#
+# Operators: copy this file to <datadir>/defaults_llm_pricing.yaml (or
+# any path configured via management.json:
+#
+#   { "AgentNetwork": { "PricingDefaultsFile": "/path/defaults_llm_pricing.yaml" } }
+#
+# ) and adjust the entries you want to change. Management re-reads the
+# file periodically (mtime poll, every minute): the live table feeds the
+# proxies' cost metering and the dashboard's model-price prefill, so
+# edits apply without a restart. Your file only needs the entries you
+# want to change — but each entry REPLACES the built-in entry for that
+# surface+model whole, so repeat the cache rates you want to keep.
+# Unknown fields and negative or non-finite rates are rejected: at
+# startup that fails boot (for an explicitly configured path); at
+# runtime the previous table is kept and a warning is logged. Deleting
+# the file reverts to the built-in defaults below.
+#
+# Top-level keys are pricing surfaces — the parser shape requests are
+# metered under: "openai" (also Azure, Mistral, and OpenAI-compatible
+# gateways), "anthropic" (also Anthropic-on-Vertex), "bedrock"
+# (normalized ids, e.g. anthropic.claude-sonnet-4-5). Model keys must be
+# the normalized id the proxy meters (version/region suffixes stripped).
+#
+# Values are USD per 1_000 tokens. Optional cache fields:
+#   cached_input_per_1k    OpenAI shape: rate for cached prompt tokens
+#                          (a SUBSET of input tokens). Absent -> cached
+#                          portion bills at input_per_1k.
+#   cache_read_per_1k      Anthropic shape: rate for cache_read tokens
+#                          (ADDITIVE to input). Absent -> input rate.
+#   cache_creation_per_1k  Anthropic shape: rate for cache_creation
+#                          tokens (ADDITIVE to input). Absent -> input
+#                          rate.
+`

management/internals/modules/agentnetwork/pricing/gen.go

@@ -0,0 +1,20 @@
+//go:build ignore
+
+// Regenerates defaults_llm_pricing.example.yaml from the compiled-in
+// default pricing table. Run via:
+//
+//	go generate ./management/internals/modules/agentnetwork/pricing
+package main
+
+import (
+	"log"
+	"os"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
+)
+
+func main() {
+	if err := os.WriteFile("defaults_llm_pricing.example.yaml", pricing.MarshalDefaultsYAML(), 0o644); err != nil {
+		log.Fatalf("write defaults_llm_pricing.example.yaml: %v", err)
+	}
+}

management/internals/modules/agentnetwork/pricing/override.go

@@ -0,0 +1,244 @@
+package pricing
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"math"
+	"os"
+	"sync"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+	"gopkg.in/yaml.v3"
+)
+
+// DefaultFileName is the basename probed under management's datadir when
+// AgentNetwork.PricingDefaultsFile doesn't configure an explicit path.
+const DefaultFileName = "defaults_llm_pricing.yaml"
+
+// ReloadInterval is the cadence at which the pricing file's mtime is
+// polled for changes.
+const ReloadInterval = time.Minute
+
+// maxFileBytes bounds the pricing file read so a misconfigured path
+// (pointed at a huge file) cannot exhaust process memory.
+const maxFileBytes = 4 << 20
+
+// pricingFile mirrors the on-disk YAML schema — the same schema the
+// proxy's retired embedded defaults_pricing.yaml used, so files written
+// for it keep working. Keys are pricing surfaces ("openai", "anthropic",
+// "bedrock"); nested keys are normalized model ids.
+type pricingFile map[string]map[string]struct {
+	InputPer1k         float64 `yaml:"input_per_1k"`
+	OutputPer1k        float64 `yaml:"output_per_1k"`
+	CachedInputPer1k   float64 `yaml:"cached_input_per_1k"`
+	CacheReadPer1k     float64 `yaml:"cache_read_per_1k"`
+	CacheCreationPer1k float64 `yaml:"cache_creation_per_1k"`
+}
+
+// fileState tracks the watched pricing file across reloads.
+var fileState struct {
+	mu    sync.Mutex
+	path  string
+	mtime int64
+}
+
+// LoadFile loads the management-side pricing defaults file and makes it
+// the live table (merged entry-whole over the compiled-in fallback; see
+// DefaultTable). The path stays registered for the periodic reloader, so
+// later edits — or the file (re)appearing after deletion — are picked up
+// without a restart.
+//
+// required governs the missing-file case: true for an explicitly
+// configured path (a typo must fail startup rather than silently bill
+// with built-ins the operator believes they replaced), false for the
+// conventional <datadir>/defaults_llm_pricing.yaml probe (absent file =
+// compiled-in defaults, still watched in case it appears). A file that
+// exists but is malformed is always an error at load time.
+func LoadFile(path string, required bool) error {
+	if path == "" {
+		return nil
+	}
+	fileState.mu.Lock()
+	fileState.path = path
+	fileState.mu.Unlock()
+
+	table, mtime, err := readFile(path)
+	if err != nil {
+		if errors.Is(err, fs.ErrNotExist) && !required {
+			return nil
+		}
+		return err
+	}
+	storeFileTable(table, mtime)
+	return nil
+}
+
+// StartReloader launches the periodic mtime-poll goroutine for the file
+// registered by LoadFile. Runtime failures are lenient — a parse error
+// keeps the previously loaded table and logs a warning; a deleted file
+// reverts to the compiled-in defaults — so a mid-edit save can never
+// take pricing down. Returns immediately when no path was registered.
+func StartReloader(ctx context.Context, interval time.Duration) {
+	fileState.mu.Lock()
+	path := fileState.path
+	fileState.mu.Unlock()
+	if path == "" {
+		return
+	}
+	if interval <= 0 {
+		interval = ReloadInterval
+	}
+	go func() {
+		t := time.NewTicker(interval)
+		defer t.Stop()
+		for {
+			select {
+			case <-ctx.Done():
+				return
+			case <-t.C:
+				reload()
+			}
+		}
+	}()
+}
+
+// reload performs one mtime check + reload cycle.
+func reload() {
+	fileState.mu.Lock()
+	path, lastMtime := fileState.path, fileState.mtime
+	fileState.mu.Unlock()
+
+	st, err := os.Stat(path)
+	if err != nil {
+		if errors.Is(err, fs.ErrNotExist) {
+			// File removed (or not yet created): serve compiled-in
+			// defaults and reset mtime so a future (re)appearance loads.
+			if mergedTable.Swap(nil) != nil {
+				log.Warnf("agent-network pricing defaults file %s removed; reverting to built-in defaults", path)
+			}
+			setMtime(0)
+			return
+		}
+		log.Warnf("agent-network pricing defaults reload: stat %s: %v", path, err)
+		return
+	}
+	if st.ModTime().UnixNano() == lastMtime {
+		return
+	}
+
+	table, mtime, err := readFile(path)
+	if err != nil {
+		// Keep the previously loaded table — never blank prices because
+		// an operator saved mid-edit.
+		log.Warnf("agent-network pricing defaults reload failed for %s (keeping previous table): %v", path, err)
+		return
+	}
+	storeFileTable(table, mtime)
+	log.Infof("agent-network pricing defaults reloaded from %s", path)
+}
+
+func readFile(path string) (map[string]map[string]Entry, int64, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, 0, fmt.Errorf("open pricing defaults %s: %w", path, err)
+	}
+	defer func() { _ = f.Close() }()
+
+	st, err := f.Stat()
+	if err != nil {
+		return nil, 0, fmt.Errorf("stat pricing defaults %s: %w", path, err)
+	}
+	data, err := io.ReadAll(io.LimitReader(f, maxFileBytes+1))
+	if err != nil {
+		return nil, 0, fmt.Errorf("read pricing defaults %s: %w", path, err)
+	}
+	if len(data) > maxFileBytes {
+		return nil, 0, fmt.Errorf("pricing defaults %s exceeds %d bytes", path, maxFileBytes)
+	}
+	table, err := parsePricingYAML(data)
+	if err != nil {
+		return nil, 0, fmt.Errorf("parse pricing defaults %s: %w", path, err)
+	}
+	return table, st.ModTime().UnixNano(), nil
+}
+
+// storeFileTable merges the parsed file over the compiled-in base and
+// publishes the result as the live table. File entries replace the
+// built-in entry for the same (surface, model) whole — they are not
+// field-merged — and surfaces/models the file doesn't mention keep the
+// built-in rates, so a partial file only needs the entries it changes.
+func storeFileTable(table map[string]map[string]Entry, mtime int64) {
+	base := compiledBase()
+	merged := make(map[string]map[string]Entry, len(base)+len(table))
+	for surface, models := range base {
+		inner := make(map[string]Entry, len(models))
+		for id, e := range models {
+			inner[id] = e
+		}
+		merged[surface] = inner
+	}
+	for surface, models := range table {
+		inner, ok := merged[surface]
+		if !ok {
+			inner = make(map[string]Entry, len(models))
+			merged[surface] = inner
+		}
+		for id, e := range models {
+			inner[id] = e
+		}
+	}
+	mergedTable.Store(&merged)
+	setMtime(mtime)
+}
+
+func setMtime(v int64) {
+	fileState.mu.Lock()
+	fileState.mtime = v
+	fileState.mu.Unlock()
+}
+
+// parsePricingYAML decodes and validates the pricing YAML. Unknown
+// fields are rejected (typos surface instead of silently pricing at 0)
+// and every rate must be a finite, non-negative USD amount — the same
+// constraints the HTTP API enforces on operator per-provider prices.
+func parsePricingYAML(data []byte) (map[string]map[string]Entry, error) {
+	dec := yaml.NewDecoder(bytes.NewReader(data))
+	dec.KnownFields(true)
+
+	var raw pricingFile
+	if err := dec.Decode(&raw); err != nil && !errors.Is(err, io.EOF) {
+		return nil, fmt.Errorf("decode yaml: %w", err)
+	}
+
+	out := make(map[string]map[string]Entry, len(raw))
+	for surface, models := range raw {
+		inner := make(map[string]Entry, len(models))
+		for model, e := range models {
+			for field, v := range map[string]float64{
+				"input_per_1k":          e.InputPer1k,
+				"output_per_1k":         e.OutputPer1k,
+				"cached_input_per_1k":   e.CachedInputPer1k,
+				"cache_read_per_1k":     e.CacheReadPer1k,
+				"cache_creation_per_1k": e.CacheCreationPer1k,
+			} {
+				if v < 0 || math.IsNaN(v) || math.IsInf(v, 0) {
+					return nil, fmt.Errorf("%s/%s: %s must be a finite, non-negative rate, got %v", surface, model, field, v)
+				}
+			}
+			inner[model] = Entry{
+				InputPer1k:         e.InputPer1k,
+				OutputPer1k:        e.OutputPer1k,
+				CachedInputPer1k:   e.CachedInputPer1k,
+				CacheReadPer1k:     e.CacheReadPer1k,
+				CacheCreationPer1k: e.CacheCreationPer1k,
+			}
+		}
+		out[surface] = inner
+	}
+	return out, nil
+}

management/internals/modules/agentnetwork/pricing/override_test.go

@@ -0,0 +1,173 @@
+package pricing
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// resetFileState snapshots and restores the package-level file state so
+// tests stay order-independent.
+func resetFileState(t *testing.T) {
+	t.Helper()
+	prevMerged := mergedTable.Load()
+	fileState.mu.Lock()
+	prevPath, prevMtime := fileState.path, fileState.mtime
+	fileState.mu.Unlock()
+	t.Cleanup(func() {
+		mergedTable.Store(prevMerged)
+		fileState.mu.Lock()
+		fileState.path, fileState.mtime = prevPath, prevMtime
+		fileState.mu.Unlock()
+	})
+}
+
+func writePricing(t *testing.T, path, yml string) {
+	t.Helper()
+	require.NoError(t, os.WriteFile(path, []byte(yml), 0o600))
+}
+
+func TestLoadFile_MergesOverCompiledDefaults(t *testing.T) {
+	resetFileState(t)
+	path := filepath.Join(t.TempDir(), DefaultFileName)
+	writePricing(t, path, `
+openai:
+  # Reprice a built-in model. The entry replaces the built-in WHOLE:
+  # omitting the cache rate here drops the built-in 0.00125 discount.
+  gpt-4o:
+    input_per_1k: 0.9
+    output_per_1k: 1.8
+  # A model NetBird doesn't know at all.
+  my-private-ft:
+    input_per_1k: 0.01
+    output_per_1k: 0.02
+    cached_input_per_1k: 0.005
+gemini:
+  gemini-pro:
+    input_per_1k: 0.00125
+    output_per_1k: 0.005
+`)
+	require.NoError(t, LoadFile(path, true))
+	table := DefaultTable()
+
+	gpt4o := table["openai"]["gpt-4o"]
+	assert.InDelta(t, 0.9, gpt4o.InputPer1k, 1e-9, "file rate replaces the compiled-in rate")
+	assert.Zero(t, gpt4o.CachedInputPer1k, "entries replace whole — omitted cache rate is dropped, not inherited")
+
+	ft := table["openai"]["my-private-ft"]
+	assert.InDelta(t, 0.005, ft.CachedInputPer1k, 1e-9, "unknown models are added to the surface")
+	_, ok := table["gemini"]["gemini-pro"]
+	assert.True(t, ok, "a surface the catalog doesn't declare can be added")
+
+	// Untouched entries keep compiled-in rates (catalog, other surface,
+	// supplemental).
+	assert.InDelta(t, 0.00015, table["openai"]["gpt-4o-mini"].InputPer1k, 1e-9, "unlisted model keeps compiled rate")
+	assert.InDelta(t, 0.003, table["anthropic"]["claude-sonnet-4-5"].InputPer1k, 1e-9, "unlisted surface untouched")
+	assert.InDelta(t, 0.00125, table["openai"]["gpt-5"].InputPer1k, 1e-9, "supplemental entries untouched")
+
+	// The synthesizer-facing lookup reads the live table too.
+	e, ok := LookupDefault([]string{"openai"}, "gpt-4o")
+	require.True(t, ok)
+	assert.InDelta(t, 0.9, e.InputPer1k, 1e-9, "LookupDefault serves the file-backed rate")
+}
+
+func TestLoadFile_MissingPath(t *testing.T) {
+	resetFileState(t)
+	missing := filepath.Join(t.TempDir(), DefaultFileName)
+
+	require.Error(t, LoadFile(missing, true),
+		"explicitly configured path that doesn't exist must fail startup")
+
+	require.NoError(t, LoadFile(missing, false),
+		"conventional datadir probe tolerates an absent file (compiled-in defaults serve)")
+	assert.Nil(t, mergedTable.Load(), "no file, no merged table")
+	fileState.mu.Lock()
+	path := fileState.path
+	fileState.mu.Unlock()
+	assert.Equal(t, missing, path, "the path stays registered so the reloader picks the file up when it appears")
+}
+
+func TestLoadFile_RejectsInvalid(t *testing.T) {
+	resetFileState(t)
+	dir := t.TempDir()
+	cases := map[string]string{
+		"unknown field (typo)": "openai:\n  gpt-4o:\n    input_per1k: 0.1\n",
+		"negative rate":        "openai:\n  gpt-4o:\n    input_per_1k: -0.1\n",
+		"non-numeric rate":     "openai:\n  gpt-4o:\n    input_per_1k: cheap\n",
+		"not a mapping":        "- just\n- a\n- list\n",
+	}
+	for name, yml := range cases {
+		path := filepath.Join(dir, DefaultFileName)
+		writePricing(t, path, yml)
+		assert.Error(t, LoadFile(path, true), "case %q must be rejected", name)
+	}
+}
+
+// TestReload_LifeCycle drives the reloader's single-shot reload through
+// its full lifecycle: file edit picked up on mtime change, a broken save
+// keeps the previous table, and file removal reverts to the compiled-in
+// defaults (then a re-created file loads again).
+func TestReload_LifeCycle(t *testing.T) {
+	resetFileState(t)
+	path := filepath.Join(t.TempDir(), DefaultFileName)
+	writePricing(t, path, "openai:\n  gpt-4o:\n    input_per_1k: 0.5\n    output_per_1k: 1\n")
+	require.NoError(t, LoadFile(path, true))
+	require.InDelta(t, 0.5, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9)
+
+	// Edit: new mtime, new rates.
+	writePricing(t, path, "openai:\n  gpt-4o:\n    input_per_1k: 0.7\n    output_per_1k: 1.4\n")
+	bumpMtime(t, path)
+	reload()
+	assert.InDelta(t, 0.7, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9, "edit must be picked up")
+
+	// Broken save: previous table survives.
+	writePricing(t, path, "openai:\n  gpt-4o:\n    input_per_1k: -1\n")
+	bumpMtime(t, path)
+	reload()
+	assert.InDelta(t, 0.7, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9,
+		"a malformed save must keep the previously loaded table, never blank prices")
+
+	// Removal: compiled-in defaults serve again.
+	require.NoError(t, os.Remove(path))
+	reload()
+	assert.InDelta(t, 0.0025, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9,
+		"file removal reverts to the compiled-in rate")
+
+	// Re-created file loads without a restart.
+	writePricing(t, path, "openai:\n  gpt-4o:\n    input_per_1k: 0.9\n    output_per_1k: 1.8\n")
+	bumpMtime(t, path)
+	reload()
+	assert.InDelta(t, 0.9, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9,
+		"a file appearing after removal (or after a missing-probe boot) must load")
+}
+
+// bumpMtime pushes the file's mtime forward past the previously recorded
+// value — timestamps can otherwise collide within the test's timescale.
+func bumpMtime(t *testing.T, path string) {
+	t.Helper()
+	st, err := os.Stat(path)
+	require.NoError(t, err)
+	next := st.ModTime().Add(2 * 1e9)
+	require.NoError(t, os.Chtimes(path, next, next))
+}
+
+// TestExampleYAML_InSyncWithBuiltins is the golden guard for
+// defaults_llm_pricing.example.yaml: the shipped example must stay
+// byte-identical to what the compiled-in table renders (catalog edits
+// require `go generate ./management/internals/modules/agentnetwork/pricing`)
+// and must round-trip through the same parser operators' files go
+// through, reproducing the compiled-in table exactly.
+func TestExampleYAML_InSyncWithBuiltins(t *testing.T) {
+	onDisk, err := os.ReadFile("defaults_llm_pricing.example.yaml")
+	require.NoError(t, err, "example file must exist next to the package")
+	require.Equal(t, string(MarshalDefaultsYAML()), string(onDisk),
+		"defaults_llm_pricing.example.yaml is stale — run: go generate ./management/internals/modules/agentnetwork/pricing")
+
+	parsed, err := parsePricingYAML(onDisk)
+	require.NoError(t, err, "the example must be a valid pricing defaults file")
+	assert.Equal(t, buildDefaultTable(), parsed,
+		"parsing the example must reproduce the compiled-in table exactly")
+}

management/internals/modules/agentnetwork/synthesizer.go

@@ -233,6 +233,11 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 		return nil, err
 	}
 
+	costMeterJSON, err := buildCostMeterConfigJSON(enabledProviders, groupIndex)
+	if err != nil {
+		return nil, err
+	}
+
 	mergedGuardrails := mergeGuardrails(enabledPolicies, guardrailsByID)
 	applyAccountCollectionControls(&mergedGuardrails, settings)
 	// The proxy guardrail is a per-provider fail-closed backstop; the
@@ -248,7 +253,7 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 	// Use the merged decision (account settings OR policy-required redaction),
 	// not the raw account flag, so a policy that mandates PII redaction is
 	// honored by the capture parsers even when the account toggle is off.
-	middlewares := buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON, mergedGuardrails.PromptCapture.RedactPii, mergedGuardrails.PromptCapture.Enabled)
+	middlewares := buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON, costMeterJSON, mergedGuardrails.PromptCapture.RedactPii, mergedGuardrails.PromptCapture.Enabled)
 
 	priv, pub, err := pickServiceSessionKeys(enabledProviders)
 	if err != nil {
@@ -700,7 +705,7 @@ func buildIdentityExtraHeaders(p *types.Provider, extras []catalog.ExtraHeader)
 // requests bound for gateways like LiteLLM that key budgets and
 // attribution off request headers. CanMutate is required so its
 // HeadersAdd / HeadersRemove pass the framework's mutation gate.
-func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON []byte, redactPii, capturePromptContent bool) []rpservice.MiddlewareConfig {
+func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON, costMeterJSON []byte, redactPii, capturePromptContent bool) []rpservice.MiddlewareConfig {
 	// Both parsers receive an explicit capture flag derived from the account's
 	// enable_prompt_collection toggle; nil/unset would default to the legacy
 	// "always emit" behavior in the middleware, which is precisely what we
@@ -769,10 +774,13 @@ func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON []byt
 			ConfigJSON: []byte("{}"),
 		},
 		{
+			// Carries the full pricing table (defaults + per-provider
+			// operator prices) so the proxy bills without an embedded
+			// price list; see buildCostMeterConfigJSON.
 			ID:         middlewareIDCostMeter,
 			Enabled:    true,
 			Slot:       rpservice.MiddlewareSlotOnResponse,
-			ConfigJSON: []byte("{}"),
+			ConfigJSON: costMeterJSON,
 		},
 		{
 			ID:         middlewareIDLLMResponseParser,

management/internals/modules/agentnetwork/synthesizer_pricing.go

@@ -0,0 +1,131 @@
+package agentnetwork
+
+import (
+	"encoding/json"
+	"fmt"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+	sharedllm "github.com/netbirdio/netbird/shared/llm"
+)
+
+// costMeterConfig is the JSON shape the proxy-side cost_meter middleware
+// expects (mirror-type pattern, same as routerConfig). The top-level
+// "pricing" wrapper is the feature-detection signal: an old proxy's config
+// struct ignores it as an unknown field, and a new proxy treats its
+// absence as "old management" (skips every cost computation and warns).
+type costMeterConfig struct {
+	Pricing *costMeterPricing `json:"pricing,omitempty"`
+}
+
+// costMeterPricing carries the full pricing table:
+//   - Defaults: surface ("openai"/"anthropic"/"bedrock") -> normalized
+//     model id -> rates. The full default table ships to every account —
+//     it is small (~10 KB) and keeps gateway-style providers (which
+//     enumerate no models) priced for every catalog model.
+//   - Providers: provider record id (matched against the
+//     llm.resolved_provider_id metadata llm_router stamps) -> normalized
+//     model id -> rates. Entries are fully materialized here at synth
+//     time — default cache rates already folded in — so the proxy does
+//     two map lookups and no merging.
+type costMeterPricing struct {
+	Defaults  map[string]map[string]pricing.Entry `json:"defaults,omitempty"`
+	Providers map[string]map[string]pricing.Entry `json:"providers,omitempty"`
+}
+
+// buildCostMeterConfigJSON assembles the cost_meter middleware config
+// from the default pricing table plus the operator's stored per-provider
+// model prices. Same orphan rule as the router: a provider no enabled
+// policy authorises is unreachable, so its prices are not shipped.
+//
+// Overlay semantics per model row:
+//   - The row's model id is normalized exactly the way the proxy's
+//     request parser normalizes the ids it meters (bedrock ARN/region/
+//     version stripping, vertex "@version" stripping), so the per-record
+//     lookup key compares equal to llm.model at billing time.
+//   - The entry starts from the default entry for that model (when one
+//     exists) to inherit cache rates the operator didn't state.
+//   - Operator input/output overlay verbatim — including an explicit 0,
+//     which prices the model as free (self-hosted / internal endpoints)
+//     rather than silently reverting to list price.
+//   - Cache-rate pointers overlay only when non-nil: nil means "inherit
+//     the default", an explicit 0 means "no discount, bill this bucket
+//     at the input rate".
+func buildCostMeterConfigJSON(providers []*types.Provider, groupIndex map[string][]string) ([]byte, error) {
+	cfg := costMeterConfig{Pricing: &costMeterPricing{
+		Defaults: pricing.DefaultTable(),
+	}}
+
+	perRecord := make(map[string]map[string]pricing.Entry)
+	for _, p := range providers {
+		if _, hasPolicy := groupIndex[p.ID]; !hasPolicy {
+			// Orphan: unreachable via the router, so unpriceable.
+			continue
+		}
+		if len(p.Models) == 0 {
+			// Gateway-style "claim every model" provider: the defaults
+			// table is its price list.
+			continue
+		}
+		entry, _ := catalog.Lookup(p.ProviderID)
+		models := make(map[string]pricing.Entry, len(p.Models))
+		for _, m := range p.Models {
+			id := normalizePricingModelID(p.ProviderID, m.ID)
+			if id == "" {
+				continue
+			}
+			if _, dup := models[id]; dup {
+				// First occurrence wins on post-normalization duplicates,
+				// matching providerModelIDs' dedup order for routing.
+				continue
+			}
+			models[id] = materializeEntry(entry.PricingSurfaces, id, m)
+		}
+		if len(models) > 0 {
+			perRecord[p.ID] = models
+		}
+	}
+	if len(perRecord) > 0 {
+		cfg.Pricing.Providers = perRecord
+	}
+
+	out, err := json.Marshal(cfg)
+	if err != nil {
+		return nil, fmt.Errorf("marshal cost_meter middleware config: %w", err)
+	}
+	return out, nil
+}
+
+// normalizePricingModelID maps an operator-entered model id onto the
+// normalized id the proxy's request parser emits as llm.model — the key
+// the cost meter looks up at billing time.
+func normalizePricingModelID(catalogProviderID, modelID string) string {
+	switch {
+	case catalog.IsBedrockPathStyle(catalogProviderID):
+		return sharedllm.NormalizeBedrockModel(modelID)
+	case catalog.IsVertexPathStyle(catalogProviderID):
+		return sharedllm.NormalizeVertexModel(modelID)
+	default:
+		return modelID
+	}
+}
+
+// materializeEntry folds the default entry for (surfaces, model) — when
+// one exists — under the operator's stored prices, producing the fully
+// materialized wire entry.
+func materializeEntry(surfaces []string, normalizedID string, m types.ProviderModel) pricing.Entry {
+	e, _ := pricing.LookupDefault(surfaces, normalizedID) // zero Entry on miss
+	e.InputPer1k = m.InputPer1k
+	e.OutputPer1k = m.OutputPer1k
+	if m.CachedInputPer1k != nil {
+		e.CachedInputPer1k = *m.CachedInputPer1k
+	}
+	if m.CacheReadPer1k != nil {
+		e.CacheReadPer1k = *m.CacheReadPer1k
+	}
+	if m.CacheCreationPer1k != nil {
+		e.CacheCreationPer1k = *m.CacheCreationPer1k
+	}
+	return e
+}

management/internals/modules/agentnetwork/synthesizer_pricing_test.go

@@ -0,0 +1,105 @@
+package agentnetwork
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+)
+
+func fptr(v float64) *float64 { return &v }
+
+func decodeCostMeterConfig(t *testing.T, raw []byte) costMeterConfig {
+	t.Helper()
+	var cfg costMeterConfig
+	require.NoError(t, json.Unmarshal(raw, &cfg), "cost meter config must round-trip")
+	require.NotNil(t, cfg.Pricing, "pricing wrapper must be present")
+	return cfg
+}
+
+// TestBuildCostMeterConfig_BedrockModelNormalization: the operator may
+// paste region-prefixed, versioned, or ARN-wrapped Bedrock ids; the
+// per-record map must be keyed by the normalized id the request parser
+// emits as llm.model, or the lookup never hits at billing time.
+func TestBuildCostMeterConfig_BedrockModelNormalization(t *testing.T) {
+	bedrock := &types.Provider{
+		ID:         "prov-bedrock",
+		ProviderID: "bedrock_api",
+		Enabled:    true,
+		Models: []types.ProviderModel{
+			{ID: "eu.anthropic.claude-sonnet-4-5-20250929-v1:0", InputPer1k: 0.0033, OutputPer1k: 0.0165},
+			// Post-normalization duplicate of the row above under a
+			// different regional spelling — first occurrence wins.
+			{ID: "us.anthropic.claude-sonnet-4-5-20250929-v1:0", InputPer1k: 9.9, OutputPer1k: 9.9},
+		},
+	}
+	raw, err := buildCostMeterConfigJSON([]*types.Provider{bedrock}, map[string][]string{"prov-bedrock": {"grp"}})
+	require.NoError(t, err)
+	cfg := decodeCostMeterConfig(t, raw)
+
+	models := cfg.Pricing.Providers["prov-bedrock"]
+	require.Len(t, models, 1, "both spellings normalize to one model; first row wins")
+	e, ok := models["anthropic.claude-sonnet-4-5"]
+	require.True(t, ok, "key must be the normalized id the parser emits, not the operator's raw spelling")
+	assert.InDelta(t, 0.0033, e.InputPer1k, 1e-9, "first row's rate wins the dedup")
+	assert.InDelta(t, 0.0003, e.CacheReadPer1k, 1e-9, "cache read inherited from the bedrock default entry")
+	assert.InDelta(t, 0.00375, e.CacheCreationPer1k, 1e-9, "cache creation inherited from the bedrock default entry")
+}
+
+// TestBuildCostMeterConfig_CacheRateNilVsZero pins the pointer semantics:
+// nil inherits the default cache rate, explicit 0 clears it (that bucket
+// bills at the input rate on the proxy).
+func TestBuildCostMeterConfig_CacheRateNilVsZero(t *testing.T) {
+	p := &types.Provider{
+		ID:         "prov-oai",
+		ProviderID: "openai_api",
+		Enabled:    true,
+		Models: []types.ProviderModel{
+			{ID: "gpt-4o", InputPer1k: 0.002, OutputPer1k: 0.008},                                    // nil → inherit 0.00125
+			{ID: "gpt-4o-mini", InputPer1k: 0.0001, OutputPer1k: 0.0005, CachedInputPer1k: fptr(0)},  // explicit 0 → no discount
+			{ID: "my-custom-ft", InputPer1k: 0.01, OutputPer1k: 0.02, CachedInputPer1k: fptr(0.005)}, // unknown model, explicit rate
+		},
+	}
+	raw, err := buildCostMeterConfigJSON([]*types.Provider{p}, map[string][]string{"prov-oai": {"grp"}})
+	require.NoError(t, err)
+	cfg := decodeCostMeterConfig(t, raw)
+	models := cfg.Pricing.Providers["prov-oai"]
+
+	assert.InDelta(t, 0.00125, models["gpt-4o"].CachedInputPer1k, 1e-9, "nil cache pointer inherits the default rate")
+	assert.Zero(t, models["gpt-4o-mini"].CachedInputPer1k, "explicit 0 overrides the default (0.000075) — bucket bills at input rate")
+	custom := models["my-custom-ft"]
+	assert.InDelta(t, 0.005, custom.CachedInputPer1k, 1e-9, "unknown model keeps the operator's explicit cache rate")
+	assert.Zero(t, custom.CacheReadPer1k, "no default to inherit for a model outside the catalog")
+}
+
+// TestBuildCostMeterConfig_OrphanAndGatewayProviders: an orphan (no
+// authorising policy) is unreachable so its prices must not ship; a
+// gateway with no model rows relies on the defaults table and gets no
+// per-record entry.
+func TestBuildCostMeterConfig_OrphanAndGatewayProviders(t *testing.T) {
+	orphan := &types.Provider{
+		ID:         "prov-orphan",
+		ProviderID: "openai_api",
+		Enabled:    true,
+		Models:     []types.ProviderModel{{ID: "gpt-4o", InputPer1k: 1, OutputPer1k: 1}},
+	}
+	gateway := &types.Provider{
+		ID:         "prov-litellm",
+		ProviderID: "litellm_proxy",
+		Enabled:    true,
+		Models:     []types.ProviderModel{},
+	}
+	raw, err := buildCostMeterConfigJSON(
+		[]*types.Provider{orphan, gateway},
+		map[string][]string{"prov-litellm": {"grp"}}, // orphan has no policy
+	)
+	require.NoError(t, err)
+	cfg := decodeCostMeterConfig(t, raw)
+
+	assert.NotContains(t, cfg.Pricing.Providers, "prov-orphan", "orphan provider prices must not ship")
+	assert.NotContains(t, cfg.Pricing.Providers, "prov-litellm", "empty-models gateway needs no per-record entry")
+	assert.NotEmpty(t, cfg.Pricing.Defaults["openai"], "defaults still ship so the gateway's catalog-model traffic is priced")
+}

management/internals/modules/agentnetwork/synthesizer_test.go

@@ -33,14 +33,17 @@ func newSynthTestSettings() *types.Settings {
 
 func newSynthTestProvider() *types.Provider {
 	return &types.Provider{
-		ID:                "prov-1",
-		AccountID:         testAccountID,
-		ProviderID:        "openai_api",
-		Name:              "OpenAI",
-		UpstreamURL:       "https://api.openai.com",
-		APIKey:            "sk-test-key",
-		Enabled:           true,
-		Models:            []types.ProviderModel{{ID: "gpt-5.4", InputPer1k: 0.0025, OutputPer1k: 0.015}},
+		ID:          "prov-1",
+		AccountID:   testAccountID,
+		ProviderID:  "openai_api",
+		Name:        "OpenAI",
+		UpstreamURL: "https://api.openai.com",
+		APIKey:      "sk-test-key",
+		Enabled:     true,
+		// Prices deliberately differ from the catalog's gpt-5.4 rates
+		// (0.0025/0.015) so pricing tests can prove the operator's
+		// stored price overlays the catalog default.
+		Models:            []types.ProviderModel{{ID: "gpt-5.4", InputPer1k: 0.004, OutputPer1k: 0.02}},
 		SessionPrivateKey: "test-priv-key",
 		SessionPublicKey:  "test-pub-key",
 		CreatedAt:         time.Date(2026, 1, 1, 0, 0, 0, 0, time.UTC),
@@ -214,7 +217,27 @@ func TestSynthesizeServices_HappyPath(t *testing.T) {
 
 	assert.Equal(t, middlewareIDCostMeter, mws[6].ID, "seventh middleware is the cost meter")
 	assert.Equal(t, rpservice.MiddlewareSlotOnResponse, mws[6].Slot, "cost meter runs on_response")
-	assert.Equal(t, []byte("{}"), mws[6].ConfigJSON, "cost meter carries an explicit empty config")
+
+	var costCfg costMeterConfig
+	require.NoError(t, json.Unmarshal(mws[6].ConfigJSON, &costCfg), "cost meter config must unmarshal")
+	require.NotNil(t, costCfg.Pricing, "cost meter config must carry the pricing table — its absence tells the proxy management predates config-delivered pricing")
+
+	gpt4o, ok := costCfg.Pricing.Defaults["openai"]["gpt-4o"]
+	require.True(t, ok, "the full default table ships regardless of the account's providers")
+	assert.InDelta(t, 0.0025, gpt4o.InputPer1k, 1e-9, "default gpt-4o input rate comes from the catalog")
+
+	openaiPrices, ok := costCfg.Pricing.Providers[openai.ID]
+	require.True(t, ok, "operator-priced provider must have a per-record entry")
+	gpt54, ok := openaiPrices["gpt-5.4"]
+	require.True(t, ok, "operator's model row keys the per-record map")
+	assert.InDelta(t, 0.004, gpt54.InputPer1k, 1e-9, "operator input price overlays the catalog default (0.0025)")
+	assert.InDelta(t, 0.02, gpt54.OutputPer1k, 1e-9, "operator output price overlays the catalog default (0.015)")
+	assert.InDelta(t, 0.00025, gpt54.CachedInputPer1k, 1e-9, "cache rate the operator didn't state is inherited from the default entry")
+
+	opus, ok := costCfg.Pricing.Providers[anthropic.ID]["claude-opus-4-7"]
+	require.True(t, ok, "anthropic's model row keys its per-record map")
+	assert.Zero(t, opus.InputPer1k, "operator-stored zero prices ship verbatim — an explicit $0 model bills as free, it does not revert to list price")
+	assert.InDelta(t, 0.0005, opus.CacheReadPer1k, 1e-9, "cache rates still inherit from the default entry")
 
 	assert.Equal(t, middlewareIDLLMResponseParser, mws[7].ID, "eighth middleware is the response parser")
 	assert.Equal(t, rpservice.MiddlewareSlotOnResponse, mws[7].Slot, "response parser runs on_response")

management/internals/modules/agentnetwork/types/provider.go

@@ -14,10 +14,24 @@ import (
 // ProviderModel is one row in the provider's models list. The operator
 // pins the per-1k input/output price for cost tracking; ID is the
 // model identifier the upstream provider expects on the wire.
+//
+// The three cache rates are pointers because absence is meaningful: nil
+// means "inherit NetBird's default rate for this model" (folded in at
+// synthesis time), while an explicit 0 means "no discount — bill this
+// cache bucket at the input rate".
 type ProviderModel struct {
 	ID          string  `json:"id"`
 	InputPer1k  float64 `json:"input_per_1k"`
 	OutputPer1k float64 `json:"output_per_1k"`
+	// CachedInputPer1k is the OpenAI-shape rate for cached prompt tokens
+	// (a subset of input tokens).
+	CachedInputPer1k *float64 `json:"cached_input_per_1k,omitempty"`
+	// CacheReadPer1k is the Anthropic-shape rate for cache-read tokens
+	// (additive to input tokens).
+	CacheReadPer1k *float64 `json:"cache_read_per_1k,omitempty"`
+	// CacheCreationPer1k is the Anthropic-shape rate for cache-creation
+	// tokens (additive to input tokens).
+	CacheCreationPer1k *float64 `json:"cache_creation_per_1k,omitempty"`
 }
 
 // Provider is an Agent Network AI provider record persisted per account.
@@ -128,9 +142,12 @@ func (p *Provider) FromAPIRequest(req *api.AgentNetworkProviderRequest) {
 	if req.Models != nil {
 		for _, m := range *req.Models {
 			p.Models = append(p.Models, ProviderModel{
-				ID:          m.Id,
-				InputPer1k:  m.InputPer1k,
-				OutputPer1k: m.OutputPer1k,
+				ID:                 m.Id,
+				InputPer1k:         m.InputPer1k,
+				OutputPer1k:        m.OutputPer1k,
+				CachedInputPer1k:   copyFloatPtr(m.CachedInputPer1k),
+				CacheReadPer1k:     copyFloatPtr(m.CacheReadPer1k),
+				CacheCreationPer1k: copyFloatPtr(m.CacheCreationPer1k),
 			})
 		}
 	}
@@ -164,9 +181,12 @@ func (p *Provider) ToAPIResponse() *api.AgentNetworkProvider {
 	models := make([]api.AgentNetworkProviderModel, 0, len(p.Models))
 	for _, m := range p.Models {
 		models = append(models, api.AgentNetworkProviderModel{
-			Id:          m.ID,
-			InputPer1k:  m.InputPer1k,
-			OutputPer1k: m.OutputPer1k,
+			Id:                 m.ID,
+			InputPer1k:         m.InputPer1k,
+			OutputPer1k:        m.OutputPer1k,
+			CachedInputPer1k:   copyFloatPtr(m.CachedInputPer1k),
+			CacheReadPer1k:     copyFloatPtr(m.CacheReadPer1k),
+			CacheCreationPer1k: copyFloatPtr(m.CacheCreationPer1k),
 		})
 	}
 	created := p.CreatedAt
@@ -201,11 +221,27 @@ func (p *Provider) ToAPIResponse() *api.AgentNetworkProvider {
 	return resp
 }
 
+// copyFloatPtr returns a fresh pointer to the same value, or nil. Keeps
+// stored models and API payloads from aliasing each other's rate fields.
+func copyFloatPtr(v *float64) *float64 {
+	if v == nil {
+		return nil
+	}
+	out := *v
+	return &out
+}
+
 // Copy returns a deep copy of the provider.
 func (p *Provider) Copy() *Provider {
 	clone := *p
 	if p.Models != nil {
-		clone.Models = append([]ProviderModel(nil), p.Models...)
+		clone.Models = make([]ProviderModel, len(p.Models))
+		for i, m := range p.Models {
+			m.CachedInputPer1k = copyFloatPtr(m.CachedInputPer1k)
+			m.CacheReadPer1k = copyFloatPtr(m.CacheReadPer1k)
+			m.CacheCreationPer1k = copyFloatPtr(m.CacheCreationPer1k)
+			clone.Models[i] = m
+		}
 	}
 	if p.ExtraValues != nil {
 		clone.ExtraValues = make(map[string]string, len(p.ExtraValues))

management/internals/modules/agentnetwork/wire_shape_test.go

@@ -103,6 +103,12 @@ func TestSynthesizedService_WireShape(t *testing.T) {
 
 	assert.Equal(t, middlewareIDCostMeter, mws[6].GetId(), "seventh middleware id")
 	assert.Equal(t, proto.MiddlewareSlot_MIDDLEWARE_SLOT_ON_RESPONSE, mws[6].GetSlot(), "cost meter slot")
+	var costCfg costMeterConfig
+	require.NoError(t, json.Unmarshal(mws[6].GetConfigJson(), &costCfg), "cost meter config JSON must decode from the wire")
+	require.NotNil(t, costCfg.Pricing, "the pricing table must travel on the wire — the proxy has no embedded price list to fall back to")
+	assert.NotEmpty(t, costCfg.Pricing.Defaults["openai"], "default table rides in every mapping")
+	assert.NotEmpty(t, costCfg.Pricing.Defaults["anthropic"], "default table covers all surfaces")
+	assert.NotEmpty(t, costCfg.Pricing.Defaults["bedrock"], "default table covers all surfaces")
 
 	assert.Equal(t, middlewareIDLLMResponseParser, mws[7].GetId(), "eighth middleware id")
 	assert.Equal(t, proto.MiddlewareSlot_MIDDLEWARE_SLOT_ON_RESPONSE, mws[7].GetSlot(), "response parser slot")

management/internals/server/config/config.go

@@ -55,6 +55,8 @@ type Config struct {
 
 	ReverseProxy ReverseProxy
 
+	AgentNetwork AgentNetwork
+
 	// disable default all-to-all policy
 	DisableDefaultPolicy bool
 
@@ -185,6 +187,24 @@ type StoreConfig struct {
 	Engine types.Engine
 }
 
+// AgentNetwork contains agent-network (LLM gateway) configuration.
+type AgentNetwork struct {
+	// PricingDefaultsFile is the path to the YAML file holding the default
+	// LLM pricing table (defaults_llm_pricing.yaml). Empty falls back to
+	// probing <Datadir>/defaults_llm_pricing.yaml; with no file present the
+	// compiled-in defaults serve. Schema: surface ("openai"/"anthropic"/
+	// "bedrock") -> model -> rates in USD per 1k tokens (input_per_1k,
+	// output_per_1k, and the optional cached_input_per_1k /
+	// cache_read_per_1k / cache_creation_per_1k). File entries replace the
+	// compiled-in entry for the same surface+model whole; everything else
+	// keeps the compiled-in rates. The file is re-read periodically (mtime
+	// poll), and the live table feeds both the synthesizer (what proxies
+	// bill with) and the dashboard's catalog endpoint (what model rows
+	// prefill with). An explicitly configured path that fails to load
+	// fails startup; runtime reload errors keep the previous table.
+	PricingDefaultsFile string
+}
+
 // ReverseProxy contains reverse proxy configuration in front of management.
 type ReverseProxy struct {
 	// TrustedHTTPProxies represents a list of trusted HTTP proxies by their IP prefixes.

management/internals/shared/grpc/components_envelope_response.go

@@ -50,7 +50,7 @@ func ToComponentSyncResponse(
 	// TODO (dmitri) consider using invariants?
 	//
 	enableSSH := computeSSHEnabledForPeer(components, peer)
-	peerConfig := toPeerConfig(peer, components.Network, dnsName, settings, httpConfig, deviceFlowConfig, enableSSH)
+	peerConfig := toPeerConfig(peer, components.Network, dnsName, settings, httpConfig, deviceFlowConfig, enableSSH, components.ForceRoutingPeerDNSResolution)
 
 	includeIPv6 := peer.SupportsIPv6() && peer.IPv6.IsValid()
 	useSourcePrefixes := peer.SupportsSourcePrefixes()

management/internals/shared/grpc/conversion.go

@@ -119,7 +119,7 @@ func toNetbirdConfig(config *nbconfig.Config, turnCredentials *Token, relayToken
 	return nbConfig
 }
 
-func toPeerConfig(peer *nbpeer.Peer, network *types.Network, dnsName string, settings *types.Settings, httpConfig *nbconfig.HttpServerConfig, deviceFlowConfig *nbconfig.DeviceAuthorizationFlow, enableSSH bool) *proto.PeerConfig {
+func toPeerConfig(peer *nbpeer.Peer, network *types.Network, dnsName string, settings *types.Settings, httpConfig *nbconfig.HttpServerConfig, deviceFlowConfig *nbconfig.DeviceAuthorizationFlow, enableSSH bool, forceRoutingPeerDNS bool) *proto.PeerConfig {
 	netmask, _ := network.Net.Mask.Size()
 	fqdn := peer.FQDN(dnsName)
 
@@ -135,7 +135,7 @@ func toPeerConfig(peer *nbpeer.Peer, network *types.Network, dnsName string, set
 		Address:                         fmt.Sprintf("%s/%d", peer.IP.String(), netmask),
 		SshConfig:                       sshConfig,
 		Fqdn:                            fqdn,
-		RoutingPeerDnsResolutionEnabled: settings.RoutingPeerDNSResolutionEnabled,
+		RoutingPeerDnsResolutionEnabled: settings.RoutingPeerDNSResolutionEnabled || peer.ProxyMeta.Embedded || forceRoutingPeerDNS,
 		LazyConnectionEnabled:           settings.LazyConnectionEnabled,
 		AutoUpdate: &proto.AutoUpdateSettings{
 			Version:      settings.AutoUpdateVersion,
@@ -162,12 +162,12 @@ func ToSyncResponse(ctx context.Context, config *nbconfig.Config, httpConfig *nb
 	useSourcePrefixes := peer.SupportsSourcePrefixes()
 
 	response := &proto.SyncResponse{
-		PeerConfig: toPeerConfig(peer, networkMap.Network, dnsName, settings, httpConfig, deviceFlowConfig, networkMap.EnableSSH),
+		PeerConfig: toPeerConfig(peer, networkMap.Network, dnsName, settings, httpConfig, deviceFlowConfig, networkMap.EnableSSH, networkMap.ForceRoutingPeerDNSResolution),
 		NetworkMap: &proto.NetworkMap{
 			Serial:     networkMap.Network.CurrentSerial(),
 			Routes:     networkmap.ToProtocolRoutes(networkMap.Routes),
 			DNSConfig:  networkmap.ToProtocolDNSConfig(networkMap.DNSConfig, dnsCache, dnsFwdPort),
-			PeerConfig: toPeerConfig(peer, networkMap.Network, dnsName, settings, httpConfig, deviceFlowConfig, networkMap.EnableSSH),
+			PeerConfig: toPeerConfig(peer, networkMap.Network, dnsName, settings, httpConfig, deviceFlowConfig, networkMap.EnableSSH, networkMap.ForceRoutingPeerDNSResolution),
 		},
 		Checks: toProtocolChecks(ctx, checks),
 	}

management/internals/shared/grpc/conversion_test.go

@@ -2,6 +2,7 @@ package grpc
 
 import (
 	"fmt"
+	"net"
 	"net/netip"
 	"reflect"
 	"testing"
@@ -14,6 +15,7 @@ import (
 	"github.com/netbirdio/netbird/management/internals/controllers/network_map"
 	"github.com/netbirdio/netbird/management/internals/controllers/network_map/controller/cache"
 	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
+	nbpeer "github.com/netbirdio/netbird/management/server/peer"
 	"github.com/netbirdio/netbird/management/server/types"
 	"github.com/netbirdio/netbird/shared/management/networkmap"
 )
@@ -301,3 +303,35 @@ func TestToNetbirdConfig_RelayInvariant(t *testing.T) {
 		assert.True(t, nbCfg.Metrics.Enabled, "metrics flag should carry the settings value")
 	})
 }
+
+func TestToPeerConfig_RoutingPeerDNSResolution(t *testing.T) {
+	network := &types.Network{Net: net.IPNet{IP: net.IPv4(100, 0, 0, 0), Mask: net.CIDRMask(8, 32)}}
+
+	newPeer := func(embedded bool) *nbpeer.Peer {
+		p := &nbpeer.Peer{IP: netip.MustParseAddr("100.0.0.1")}
+		p.ProxyMeta.Embedded = embedded
+		return p
+	}
+
+	tests := []struct {
+		name        string
+		globalFlag  bool
+		embedded    bool
+		forceParam  bool
+		wantEnabled bool
+	}{
+		{name: "global off, regular peer, no force", wantEnabled: false},
+		{name: "global on wins", globalFlag: true, wantEnabled: true},
+		{name: "embedded proxy peer forced", embedded: true, wantEnabled: true},
+		{name: "routing peer forced via param", forceParam: true, wantEnabled: true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			settings := &types.Settings{RoutingPeerDNSResolutionEnabled: tt.globalFlag}
+			cfg := toPeerConfig(newPeer(tt.embedded), network, "netbird.selfhosted", settings, nil, nil, false, tt.forceParam)
+			assert.Equal(t, tt.wantEnabled, cfg.RoutingPeerDnsResolutionEnabled,
+				"RoutingPeerDnsResolutionEnabled should reflect global || embedded || forced")
+		})
+	}
+}

management/internals/shared/grpc/server.go

@@ -921,7 +921,7 @@ func (s *Server) prepareLoginResponse(ctx context.Context, peer *nbpeer.Peer, ne
 	// if peer has reached this point then it has logged in
 	loginResp := &proto.LoginResponse{
 		NetbirdConfig: toNetbirdConfig(s.config, nil, relayToken, nil, settings),
-		PeerConfig:    toPeerConfig(peer, network, s.networkMapController.GetDNSDomain(settings), settings, s.config.HttpConfig, s.config.DeviceAuthorizationFlow, enableSSH),
+		PeerConfig:    toPeerConfig(peer, network, s.networkMapController.GetDNSDomain(settings), settings, s.config.HttpConfig, s.config.DeviceAuthorizationFlow, enableSSH, false),
 		Checks:        toProtocolChecks(ctx, postureChecks),
 	}
 

management/server/store/sql_store.go

@@ -6,6 +6,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"math"
 	"net"
 	"net/netip"
 	"net/url"
@@ -2258,117 +2259,30 @@ func (s *SqlStore) getPostureChecks(ctx context.Context, accountID string) ([]*p
 	return checks, nil
 }
 
-func (s *SqlStore) getServices(ctx context.Context, accountID string) ([]*rpservice.Service, error) {
-	const serviceQuery = `SELECT id, account_id, name, domain, enabled, auth,
-		meta_created_at, meta_certificate_issued_at, meta_status, proxy_cluster,
-		pass_host_header, rewrite_redirects, session_private_key, session_public_key,
-		mode, listen_port, port_auto_assigned, source, source_peer, terminated,
-		private, access_groups
-		FROM services WHERE account_id = $1`
+// serviceSelectColumns and targetSelectColumns are the column lists the Postgres
+// pgx read path scans. They must stay in sync with the rpservice.Service and
+// rpservice.Target gorm models; TestPgxServiceColumnsMatchGorm enforces this.
+const serviceSelectColumns = `id, account_id, name, domain, enabled, auth, restrictions,
+	meta_created_at, meta_certificate_issued_at, meta_last_renewed_at, meta_status, proxy_cluster,
+	pass_host_header, rewrite_redirects, session_private_key, session_public_key,
+	mode, listen_port, port_auto_assigned, source, source_peer, terminated,
+	private, access_groups`
 
-	const targetsQuery = `SELECT id, account_id, service_id, path, host, port, protocol,
-		target_id, target_type, enabled
-		FROM targets WHERE service_id = ANY($1)`
+const targetSelectColumns = `id, account_id, service_id, path, host, port, protocol,
+	target_id, target_type, enabled, proxy_protocol,
+	skip_tls_verify, request_timeout, session_idle_timeout, path_rewrite, custom_headers,
+	direct_upstream, middlewares, capture_max_request_bytes, capture_max_response_bytes,
+	capture_content_types, agent_network, disable_access_log`
+
+func (s *SqlStore) getServices(ctx context.Context, accountID string) ([]*rpservice.Service, error) {
+	const serviceQuery = `SELECT ` + serviceSelectColumns + ` FROM services WHERE account_id = $1`
 
 	serviceRows, err := s.pool.Query(ctx, serviceQuery, accountID)
 	if err != nil {
 		return nil, err
 	}
 
-	services, err := pgx.CollectRows(serviceRows, func(row pgx.CollectableRow) (*rpservice.Service, error) {
-		var s rpservice.Service
-		var auth []byte
-		var accessGroups []byte
-		var createdAt, certIssuedAt sql.NullTime
-		var status, proxyCluster, sessionPrivateKey, sessionPublicKey sql.NullString
-		var mode, source, sourcePeer sql.NullString
-		var terminated, portAutoAssigned, private sql.NullBool
-		var listenPort sql.NullInt64
-		err := row.Scan(
-			&s.ID,
-			&s.AccountID,
-			&s.Name,
-			&s.Domain,
-			&s.Enabled,
-			&auth,
-			&createdAt,
-			&certIssuedAt,
-			&status,
-			&proxyCluster,
-			&s.PassHostHeader,
-			&s.RewriteRedirects,
-			&sessionPrivateKey,
-			&sessionPublicKey,
-			&mode,
-			&listenPort,
-			&portAutoAssigned,
-			&source,
-			&sourcePeer,
-			&terminated,
-			&private,
-			&accessGroups,
-		)
-		if err != nil {
-			return nil, err
-		}
-
-		if auth != nil {
-			if err := json.Unmarshal(auth, &s.Auth); err != nil {
-				return nil, err
-			}
-		}
-
-		if len(accessGroups) > 0 {
-			if err := json.Unmarshal(accessGroups, &s.AccessGroups); err != nil {
-				return nil, fmt.Errorf("unmarshal access_groups: %w", err)
-			}
-		}
-
-		if private.Valid {
-			s.Private = private.Bool
-		}
-
-		s.Meta = rpservice.Meta{}
-		if createdAt.Valid {
-			s.Meta.CreatedAt = createdAt.Time
-		}
-		if certIssuedAt.Valid {
-			t := certIssuedAt.Time
-			s.Meta.CertificateIssuedAt = &t
-		}
-		if status.Valid {
-			s.Meta.Status = status.String
-		}
-		if proxyCluster.Valid {
-			s.ProxyCluster = proxyCluster.String
-		}
-		if sessionPrivateKey.Valid {
-			s.SessionPrivateKey = sessionPrivateKey.String
-		}
-		if sessionPublicKey.Valid {
-			s.SessionPublicKey = sessionPublicKey.String
-		}
-		if mode.Valid {
-			s.Mode = mode.String
-		}
-		if source.Valid {
-			s.Source = source.String
-		}
-		if sourcePeer.Valid {
-			s.SourcePeer = sourcePeer.String
-		}
-		if terminated.Valid {
-			s.Terminated = terminated.Bool
-		}
-		if portAutoAssigned.Valid {
-			s.PortAutoAssigned = portAutoAssigned.Bool
-		}
-		if listenPort.Valid {
-			s.ListenPort = uint16(listenPort.Int64)
-		}
-		s.Targets = []*rpservice.Target{}
-		return &s, nil
-	})
+	services, err := pgx.CollectRows(serviceRows, scanService)
 	if err != nil {
 		return nil, err
 	}
@@ -2379,39 +2293,12 @@ func (s *SqlStore) getServices(ctx context.Context, accountID string) ([]*rpserv
 
 	serviceIDs := make([]string, len(services))
 	serviceMap := make(map[string]*rpservice.Service)
-	for i, s := range services {
-		serviceIDs[i] = s.ID
-		serviceMap[s.ID] = s
+	for i, svc := range services {
+		serviceIDs[i] = svc.ID
+		serviceMap[svc.ID] = svc
 	}
 
-	targetRows, err := s.pool.Query(ctx, targetsQuery, serviceIDs)
-	if err != nil {
-		return nil, err
-	}
-
-	targets, err := pgx.CollectRows(targetRows, func(row pgx.CollectableRow) (*rpservice.Target, error) {
-		var t rpservice.Target
-		var path sql.NullString
-		err := row.Scan(
-			&t.ID,
-			&t.AccountID,
-			&t.ServiceID,
-			&path,
-			&t.Host,
-			&t.Port,
-			&t.Protocol,
-			&t.TargetId,
-			&t.TargetType,
-			&t.Enabled,
-		)
-		if err != nil {
-			return nil, err
-		}
-		if path.Valid {
-			t.Path = &path.String
-		}
-		return &t, nil
-	})
+	targets, err := s.getServiceTargets(ctx, serviceIDs)
 	if err != nil {
 		return nil, err
 	}
@@ -2425,6 +2312,201 @@ func (s *SqlStore) getServices(ctx context.Context, accountID string) ([]*rpserv
 	return services, nil
 }
 
+func scanService(row pgx.CollectableRow) (*rpservice.Service, error) {
+	var s rpservice.Service
+	var auth []byte
+	var restrictions []byte
+	var accessGroups []byte
+	var createdAt, certIssuedAt, lastRenewedAt sql.NullTime
+	var status, proxyCluster, sessionPrivateKey, sessionPublicKey sql.NullString
+	var mode, source, sourcePeer sql.NullString
+	var terminated, portAutoAssigned, private sql.NullBool
+	var listenPort sql.NullInt64
+	err := row.Scan(
+		&s.ID,
+		&s.AccountID,
+		&s.Name,
+		&s.Domain,
+		&s.Enabled,
+		&auth,
+		&restrictions,
+		&createdAt,
+		&certIssuedAt,
+		&lastRenewedAt,
+		&status,
+		&proxyCluster,
+		&s.PassHostHeader,
+		&s.RewriteRedirects,
+		&sessionPrivateKey,
+		&sessionPublicKey,
+		&mode,
+		&listenPort,
+		&portAutoAssigned,
+		&source,
+		&sourcePeer,
+		&terminated,
+		&private,
+		&accessGroups,
+	)
+	if err != nil {
+		return nil, err
+	}
+
+	if auth != nil {
+		if err := json.Unmarshal(auth, &s.Auth); err != nil {
+			return nil, err
+		}
+	}
+
+	if len(restrictions) > 0 {
+		if err := json.Unmarshal(restrictions, &s.Restrictions); err != nil {
+			return nil, fmt.Errorf("unmarshal restrictions: %w", err)
+		}
+	}
+
+	if len(accessGroups) > 0 {
+		if err := json.Unmarshal(accessGroups, &s.AccessGroups); err != nil {
+			return nil, fmt.Errorf("unmarshal access_groups: %w", err)
+		}
+	}
+
+	if private.Valid {
+		s.Private = private.Bool
+	}
+
+	s.Meta = serviceMetaFromRow(createdAt, certIssuedAt, lastRenewedAt, status)
+	if proxyCluster.Valid {
+		s.ProxyCluster = proxyCluster.String
+	}
+	if sessionPrivateKey.Valid {
+		s.SessionPrivateKey = sessionPrivateKey.String
+	}
+	if sessionPublicKey.Valid {
+		s.SessionPublicKey = sessionPublicKey.String
+	}
+	if mode.Valid {
+		s.Mode = mode.String
+	}
+	if source.Valid {
+		s.Source = source.String
+	}
+	if sourcePeer.Valid {
+		s.SourcePeer = sourcePeer.String
+	}
+	if terminated.Valid {
+		s.Terminated = terminated.Bool
+	}
+	if portAutoAssigned.Valid {
+		s.PortAutoAssigned = portAutoAssigned.Bool
+	}
+	if listenPort.Valid {
+		if listenPort.Int64 < 0 || listenPort.Int64 > math.MaxUint16 {
+			return nil, fmt.Errorf("listen_port %d out of range", listenPort.Int64)
+		}
+		s.ListenPort = uint16(listenPort.Int64)
+	}
+	s.Targets = []*rpservice.Target{}
+	return &s, nil
+}
+
+func serviceMetaFromRow(createdAt, certIssuedAt, lastRenewedAt sql.NullTime, status sql.NullString) rpservice.Meta {
+	meta := rpservice.Meta{}
+	if createdAt.Valid {
+		meta.CreatedAt = createdAt.Time
+	}
+	if certIssuedAt.Valid {
+		t := certIssuedAt.Time
+		meta.CertificateIssuedAt = &t
+	}
+	if lastRenewedAt.Valid {
+		t := lastRenewedAt.Time
+		meta.LastRenewedAt = &t
+	}
+	if status.Valid {
+		meta.Status = status.String
+	}
+	return meta
+}
+
+func (s *SqlStore) getServiceTargets(ctx context.Context, serviceIDs []string) ([]*rpservice.Target, error) {
+	const targetsQuery = `SELECT ` + targetSelectColumns + ` FROM targets WHERE service_id = ANY($1)`
+
+	rows, err := s.pool.Query(ctx, targetsQuery, serviceIDs)
+	if err != nil {
+		return nil, err
+	}
+
+	return pgx.CollectRows(rows, scanTarget)
+}
+
+func scanTarget(row pgx.CollectableRow) (*rpservice.Target, error) {
+	var t rpservice.Target
+	var path sql.NullString
+	var pathRewrite sql.NullString
+	var proxyProtocol, skipTLSVerify, directUpstream, agentNetwork, disableAccessLog sql.NullBool
+	var requestTimeout, sessionIdleTimeout, captureMaxRequestBytes, captureMaxResponseBytes sql.NullInt64
+	var customHeaders, middlewares, captureContentTypes []byte
+	err := row.Scan(
+		&t.ID,
+		&t.AccountID,
+		&t.ServiceID,
+		&path,
+		&t.Host,
+		&t.Port,
+		&t.Protocol,
+		&t.TargetId,
+		&t.TargetType,
+		&t.Enabled,
+		&proxyProtocol,
+		&skipTLSVerify,
+		&requestTimeout,
+		&sessionIdleTimeout,
+		&pathRewrite,
+		&customHeaders,
+		&directUpstream,
+		&middlewares,
+		&captureMaxRequestBytes,
+		&captureMaxResponseBytes,
+		&captureContentTypes,
+		&agentNetwork,
+		&disableAccessLog,
+	)
+	if err != nil {
+		return nil, err
+	}
+	if path.Valid {
+		t.Path = &path.String
+	}
+
+	t.ProxyProtocol = proxyProtocol.Bool
+	t.Options.SkipTLSVerify = skipTLSVerify.Bool
+	t.Options.RequestTimeout = time.Duration(requestTimeout.Int64)
+	t.Options.SessionIdleTimeout = time.Duration(sessionIdleTimeout.Int64)
+	t.Options.PathRewrite = rpservice.PathRewriteMode(pathRewrite.String)
+	t.Options.DirectUpstream = directUpstream.Bool
+	t.Options.CaptureMaxRequestBytes = captureMaxRequestBytes.Int64
+	t.Options.CaptureMaxResponseBytes = captureMaxResponseBytes.Int64
+	t.Options.AgentNetwork = agentNetwork.Bool
+	t.Options.DisableAccessLog = disableAccessLog.Bool
+
+	if len(customHeaders) > 0 {
+		if err := json.Unmarshal(customHeaders, &t.Options.CustomHeaders); err != nil {
+			return nil, fmt.Errorf("unmarshal custom_headers: %w", err)
+		}
+	}
+	if len(middlewares) > 0 {
+		if err := json.Unmarshal(middlewares, &t.Options.Middlewares); err != nil {
+			return nil, fmt.Errorf("unmarshal middlewares: %w", err)
+		}
+	}
+	if len(captureContentTypes) > 0 {
+		if err := json.Unmarshal(captureContentTypes, &t.Options.CaptureContentTypes); err != nil {
+			return nil, fmt.Errorf("unmarshal capture_content_types: %w", err)
+		}
+	}
+	return &t, nil
+}
+
 func (s *SqlStore) getNetworks(ctx context.Context, accountID string) ([]*networkTypes.Network, error) {
 	const query = `SELECT id, account_id, public_id, name, description FROM networks WHERE account_id = $1`
 	rows, err := s.pool.Query(ctx, query, accountID)

management/server/store/sql_store_pgx_parity_test.go

@@ -0,0 +1,74 @@
+package store
+
+import (
+	"strings"
+	"sync"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"gorm.io/gorm/schema"
+
+	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
+)
+
+// TestPgxServiceColumnsMatchGorm guards the Postgres pgx read path against
+// drifting from the gorm model. The SQLite/MySQL gorm path loads rows by struct,
+// so a new column on a model is picked up automatically, but the hand-written
+// pgx SELECT in sql_store.go must be updated by hand. This test fails when a
+// gorm column is missing from the pgx column list, which otherwise silently
+// returns zero-valued on Postgres with no compile error.
+func TestPgxServiceColumnsMatchGorm(t *testing.T) {
+	tests := []struct {
+		name          string
+		model         any
+		selectColumns string
+		// excluded lists gorm columns intentionally not loaded by the pgx path.
+		excluded map[string]struct{}
+	}{
+		{
+			name:          "service",
+			model:         &rpservice.Service{},
+			selectColumns: serviceSelectColumns,
+		},
+		{
+			name:          "target",
+			model:         &rpservice.Target{},
+			selectColumns: targetSelectColumns,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			selected := parseColumnList(tc.selectColumns)
+			for _, col := range gormColumnNames(t, tc.model) {
+				if _, ok := tc.excluded[col]; ok {
+					continue
+				}
+				_, ok := selected[col]
+				assert.Truef(t, ok,
+					"gorm column %q is not read by the Postgres pgx SELECT; add it to %sSelectColumns in sql_store.go (or to the test's excluded set if it is intentionally not loaded)",
+					col, tc.name)
+			}
+		})
+	}
+}
+
+func parseColumnList(cols string) map[string]struct{} {
+	set := make(map[string]struct{})
+	for _, c := range strings.Split(cols, ",") {
+		if c = strings.TrimSpace(c); c != "" {
+			set[c] = struct{}{}
+		}
+	}
+	return set
+}
+
+// gormColumnNames returns the DB column names gorm would migrate for the model,
+// using the same default naming strategy the store configures.
+func gormColumnNames(t *testing.T, model any) []string {
+	t.Helper()
+	sch, err := schema.Parse(model, &sync.Map{}, schema.NamingStrategy{})
+	require.NoError(t, err)
+	return sch.DBNames
+}

management/server/store/sql_store_service_test.go

@@ -5,6 +5,7 @@ import (
 	"os"
 	"runtime"
 	"testing"
+	"time"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -44,3 +45,91 @@ func TestSqlStore_GetAccount_PrivateServiceRoundtrip(t *testing.T) {
 		assert.Equal(t, []string{"grp-admins", "grp-ops"}, got.AccessGroups)
 	})
 }
+
+// TestSqlStore_GetAccount_ServiceTargetOptionsRoundtrip guards the Postgres pgx
+// read path (getServices) against silently dropping columns present on the gorm
+// model. Before the fix these fields loaded correctly on SQLite but came back
+// zero-valued on Postgres because the hand-written SELECT and scan omitted them.
+func TestSqlStore_GetAccount_ServiceTargetOptionsRoundtrip(t *testing.T) {
+	if os.Getenv("CI") == "true" && (runtime.GOOS == "darwin" || runtime.GOOS == "windows") {
+		t.Skip("skip CI tests on darwin and windows")
+	}
+
+	runTestForAllEngines(t, "", func(t *testing.T, store Store) {
+		ctx := context.Background()
+		account := newAccountWithId(ctx, "account_svc_opts", "testuser", "")
+		require.NoError(t, store.SaveAccount(ctx, account))
+
+		renewedAt := time.Now().UTC().Truncate(time.Second)
+		targetPath := "/api"
+		svc := &rpservice.Service{
+			ID:        "svc-opts",
+			AccountID: account.Id,
+			Name:      "opts-svc",
+			Domain:    "opts.example",
+			Enabled:   true,
+			Mode:      rpservice.ModeHTTP,
+			Restrictions: rpservice.AccessRestrictions{
+				AllowedCIDRs:     []string{"10.0.0.0/8"},
+				BlockedCountries: []string{"XX"},
+				CrowdSecMode:     "block",
+			},
+			Meta: rpservice.Meta{
+				LastRenewedAt: &renewedAt,
+			},
+			Targets: []*rpservice.Target{
+				{
+					AccountID:     account.Id,
+					ServiceID:     "svc-opts",
+					Path:          &targetPath,
+					Host:          "backend.internal",
+					Port:          8080,
+					Protocol:      "http",
+					TargetId:      "tgt-1",
+					Enabled:       true,
+					ProxyProtocol: true,
+					Options: rpservice.TargetOptions{
+						SkipTLSVerify:           true,
+						RequestTimeout:          30 * time.Second,
+						SessionIdleTimeout:      5 * time.Minute,
+						PathRewrite:             rpservice.PathRewritePreserve,
+						CustomHeaders:           map[string]string{"X-Foo": "bar"},
+						DirectUpstream:          true,
+						CaptureMaxRequestBytes:  1024,
+						CaptureMaxResponseBytes: 2048,
+						CaptureContentTypes:     []string{"application/json"},
+						AgentNetwork:            true,
+						DisableAccessLog:        true,
+					},
+				},
+			},
+		}
+		require.NoError(t, store.CreateService(ctx, svc))
+
+		loaded, err := store.GetAccount(ctx, account.Id)
+		require.NoError(t, err)
+		require.Len(t, loaded.Services, 1)
+
+		got := loaded.Services[0]
+		assert.Equal(t, []string{"10.0.0.0/8"}, got.Restrictions.AllowedCIDRs, "restrictions allowed CIDRs")
+		assert.Equal(t, []string{"XX"}, got.Restrictions.BlockedCountries, "restrictions blocked countries")
+		assert.Equal(t, "block", got.Restrictions.CrowdSecMode, "restrictions crowdsec mode")
+		require.NotNil(t, got.Meta.LastRenewedAt, "meta last renewed at")
+		assert.WithinDuration(t, renewedAt, *got.Meta.LastRenewedAt, time.Second, "meta last renewed at")
+
+		require.Len(t, got.Targets, 1)
+		tg := got.Targets[0]
+		assert.True(t, tg.ProxyProtocol, "target proxy protocol")
+		assert.True(t, tg.Options.SkipTLSVerify, "options skip TLS verify")
+		assert.Equal(t, 30*time.Second, tg.Options.RequestTimeout, "options request timeout")
+		assert.Equal(t, 5*time.Minute, tg.Options.SessionIdleTimeout, "options session idle timeout")
+		assert.Equal(t, rpservice.PathRewritePreserve, tg.Options.PathRewrite, "options path rewrite")
+		assert.Equal(t, map[string]string{"X-Foo": "bar"}, tg.Options.CustomHeaders, "options custom headers")
+		assert.True(t, tg.Options.DirectUpstream, "options direct upstream")
+		assert.Equal(t, int64(1024), tg.Options.CaptureMaxRequestBytes, "options capture max request bytes")
+		assert.Equal(t, int64(2048), tg.Options.CaptureMaxResponseBytes, "options capture max response bytes")
+		assert.Equal(t, []string{"application/json"}, tg.Options.CaptureContentTypes, "options capture content types")
+		assert.True(t, tg.Options.AgentNetwork, "options agent network")
+		assert.True(t, tg.Options.DisableAccessLog, "options disable access log")
+	})
+}

management/server/types/account.go

@@ -1517,6 +1517,54 @@ func (a *Account) GetResourceRoutersMap() map[string]map[string]*routerTypes.Net
 	return routers
 }
 
+// forcesRoutingPeerDNSResolution reports whether the given peer must run
+// routing-peer DNS resolution regardless of the account-global
+// RoutingPeerDNSResolutionEnabled setting. It returns true when the peer is a
+// router for a domain network resource that is targeted by an enabled
+// reverse-proxy service, so the peer's DNS forwarder starts and can resolve
+// the target for the embedded proxy peers. Embedded proxy peers themselves are
+// handled at PeerConfig build time.
+func (a *Account) forcesRoutingPeerDNSResolution(peerID string, routers map[string]map[string]*routerTypes.NetworkRouter) bool {
+	targeted := a.proxyTargetedDomainResourceIDs()
+	if len(targeted) == 0 {
+		return false
+	}
+
+	for _, resource := range a.NetworkResources {
+		if resource == nil || !resource.Enabled || resource.Type != resourceTypes.Domain {
+			continue
+		}
+		if _, ok := targeted[resource.ID]; !ok {
+			continue
+		}
+		if _, isRouter := routers[resource.NetworkID][peerID]; isRouter {
+			return true
+		}
+	}
+
+	return false
+}
+
+// proxyTargetedDomainResourceIDs returns the set of domain network resource IDs
+// targeted by an enabled, non-terminated reverse-proxy service.
+func (a *Account) proxyTargetedDomainResourceIDs() map[string]struct{} {
+	ids := make(map[string]struct{})
+	for _, svc := range a.Services {
+		if svc == nil || !svc.Enabled || svc.Terminated {
+			continue
+		}
+		for _, target := range svc.Targets {
+			if target == nil || !target.Enabled {
+				continue
+			}
+			if target.TargetType == service.TargetTypeDomain {
+				ids[target.TargetId] = struct{}{}
+			}
+		}
+	}
+	return ids
+}
+
 // getPoliciesSourcePeers collects all unique peers from the source groups defined in the given policies.
 func getPoliciesSourcePeers(policies []*Policy, groups map[string]*Group) map[string]struct{} {
 	sourcePeers := make(map[string]struct{})

management/server/types/account_components.go

@@ -140,6 +140,8 @@ func (a *Account) GetPeerNetworkMapComponents(
 		RouterPeers:               make(map[string]*ComponentPeer),
 		NetworkXIDToPublicID:      make(map[string]string, len(a.Networks)),
 		PostureCheckXIDToPublicID: make(map[string]string, len(a.PostureChecks)),
+
+		ForceRoutingPeerDNSResolution: a.forcesRoutingPeerDNSResolution(peerID, routers),
 	}
 	for _, n := range a.Networks {
 		if n != nil {

management/server/types/account_test.go

@@ -1751,3 +1751,71 @@ func hasPrivateAccessPolicy(account *Account, serviceID string) bool {
 	}
 	return false
 }
+
+func TestForcesRoutingPeerDNSResolution(t *testing.T) {
+	buildAccountRes := func(serviceEnabled, targetEnabled, resourceEnabled bool, targetType service.TargetType, resType resourceTypes.NetworkResourceType) *Account {
+		return &Account{
+			Id: "accountID",
+			Groups: map[string]*Group{
+				"router-group": {ID: "router-group", Peers: []string{"router-peer-grp"}},
+			},
+			NetworkRouters: []*routerTypes.NetworkRouter{
+				{ID: "r1", NetworkID: "net-1", AccountID: "accountID", Peer: "router-peer", Enabled: true},
+				{ID: "r2", NetworkID: "net-1", AccountID: "accountID", PeerGroups: []string{"router-group"}, Enabled: true},
+			},
+			NetworkResources: []*resourceTypes.NetworkResource{
+				{ID: "res-domain", AccountID: "accountID", NetworkID: "net-1", Type: resType, Domain: "example.org", Enabled: resourceEnabled},
+			},
+			Services: []*service.Service{
+				{
+					ID: "svc-1", AccountID: "accountID", Enabled: serviceEnabled,
+					Targets: []*service.Target{
+						{TargetId: "res-domain", TargetType: targetType, Enabled: targetEnabled},
+					},
+				},
+			},
+		}
+	}
+
+	buildAccount := func(serviceEnabled, targetEnabled, resourceEnabled bool, targetType service.TargetType) *Account {
+		return buildAccountRes(serviceEnabled, targetEnabled, resourceEnabled, targetType, resourceTypes.Domain)
+	}
+
+	t.Run("router peer for RP-targeted domain resource is forced", func(t *testing.T) {
+		account := buildAccount(true, true, true, service.TargetTypeDomain)
+		routers := account.GetResourceRoutersMap()
+		assert.True(t, account.forcesRoutingPeerDNSResolution("router-peer", routers), "direct router peer should be forced")
+		assert.True(t, account.forcesRoutingPeerDNSResolution("router-peer-grp", routers), "group-member router peer should be forced")
+	})
+
+	t.Run("non-router peer is not forced", func(t *testing.T) {
+		account := buildAccount(true, true, true, service.TargetTypeDomain)
+		assert.False(t, account.forcesRoutingPeerDNSResolution("other-peer", account.GetResourceRoutersMap()))
+	})
+
+	t.Run("not forced when service disabled", func(t *testing.T) {
+		account := buildAccount(false, true, true, service.TargetTypeDomain)
+		assert.False(t, account.forcesRoutingPeerDNSResolution("router-peer", account.GetResourceRoutersMap()))
+	})
+
+	t.Run("not forced when target disabled", func(t *testing.T) {
+		account := buildAccount(true, false, true, service.TargetTypeDomain)
+		assert.False(t, account.forcesRoutingPeerDNSResolution("router-peer", account.GetResourceRoutersMap()))
+	})
+
+	t.Run("not forced when resource disabled", func(t *testing.T) {
+		account := buildAccount(true, true, false, service.TargetTypeDomain)
+		assert.False(t, account.forcesRoutingPeerDNSResolution("router-peer", account.GetResourceRoutersMap()))
+	})
+
+	t.Run("not forced for non-domain target type", func(t *testing.T) {
+		account := buildAccount(true, true, true, service.TargetTypePeer)
+		assert.False(t, account.forcesRoutingPeerDNSResolution("router-peer", account.GetResourceRoutersMap()))
+	})
+
+	t.Run("not forced when targeted resource is not a domain", func(t *testing.T) {
+		account := buildAccountRes(true, true, true, service.TargetTypeDomain, resourceTypes.Host)
+		assert.False(t, account.forcesRoutingPeerDNSResolution("router-peer", account.GetResourceRoutersMap()),
+			"a domain target pointing at a non-domain resource must not force resolution")
+	})
+}

management/server/user.go

@@ -321,6 +321,10 @@ func (am *DefaultAccountManager) DeleteUser(ctx context.Context, accountID, init
 		return err
 	}
 
+	if targetUser.AccountID != accountID {
+		return status.NewUserNotFoundError(targetUserID)
+	}
+
 	if targetUser.Role == types.UserRoleOwner {
 		return status.NewOwnerDeletePermissionError()
 	}

management/server/user_test.go

@@ -802,6 +802,52 @@ func TestUser_DeleteUser_SelfDelete(t *testing.T) {
 	}
 }
 
+func TestUser_DeleteUser_OtherAccount(t *testing.T) {
+	testStore, cleanup, err := store.NewTestStoreFromSQL(context.Background(), "", t.TempDir())
+	if err != nil {
+		t.Fatalf("Error when creating store: %s", err)
+	}
+	t.Cleanup(cleanup)
+
+	account := newAccountWithId(context.Background(), mockAccountID, mockUserID, "", "", "", false)
+	if err = testStore.SaveAccount(context.Background(), account); err != nil {
+		t.Fatalf("Error when saving account: %s", err)
+	}
+
+	otherAccount := newAccountWithId(context.Background(), "otherAccount", "otherOwner", "", "", "", false)
+	otherAccount.Users["otherRegularUser"] = &types.User{
+		Id:        "otherRegularUser",
+		AccountID: "otherAccount",
+		Role:      types.UserRoleUser,
+	}
+	otherAccount.Users["otherServiceUser"] = &types.User{
+		Id:              "otherServiceUser",
+		AccountID:       "otherAccount",
+		Role:            types.UserRoleUser,
+		IsServiceUser:   true,
+		ServiceUserName: "otherServiceUser",
+	}
+	if err = testStore.SaveAccount(context.Background(), otherAccount); err != nil {
+		t.Fatalf("Error when saving other account: %s", err)
+	}
+
+	am := DefaultAccountManager{
+		Store:              testStore,
+		eventStore:         &activity.InMemoryEventStore{},
+		permissionsManager: permissions.NewManager(testStore),
+	}
+
+	for _, targetUserID := range []string{"otherRegularUser", "otherServiceUser"} {
+		t.Run(targetUserID, func(t *testing.T) {
+			err := am.DeleteUser(context.Background(), mockAccountID, mockUserID, targetUserID)
+			assert.Equal(t, status.NewUserNotFoundError(targetUserID), err)
+
+			_, err = testStore.GetUserByUserID(context.Background(), store.LockingStrengthNone, targetUserID)
+			assert.NoError(t, err, "user of another account must not be deleted")
+		})
+	}
+}
+
 func TestUser_DeleteUser_regularUser(t *testing.T) {
 	store, cleanup, err := store.NewTestStoreFromSQL(context.Background(), "", t.TempDir())
 	if err != nil {

proxy/internal/llm/bedrock_model.go

@@ -1,38 +0,0 @@
-package llm
-
-import (
-	"regexp"
-	"strings"
-)
-
-// bedrockRegionPrefixes are the cross-region inference-profile prefixes that
-// front a Bedrock model id (e.g. "eu.anthropic.claude-...").
-var bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
-
-// bedrockVersionSuffix matches the trailing "-vN[:N]" or "-YYYYMMDD-vN[:N]"
-// version/throughput suffix of a Bedrock model id.
-var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
-
-// NormalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
-// prefix, and the version/throughput suffix from a Bedrock model id so it
-// matches the catalog/pricing key, e.g.
-// "eu.anthropic.claude-sonnet-4-5-20250929-v1:0" -> "anthropic.claude-sonnet-4-5"
-// and the inference-profile ARN's last segment likewise. It is the single
-// source of truth shared by the request parser (which normalizes the request
-// model from the URL path) and the router (which normalizes the operator's
-// registered Bedrock model ids so both sides compare equal).
-func NormalizeBedrockModel(modelID string) string {
-	m := modelID
-	if strings.HasPrefix(m, "arn:") {
-		if i := strings.LastIndex(m, "/"); i >= 0 {
-			m = m[i+1:]
-		}
-	}
-	for _, p := range bedrockRegionPrefixes {
-		if strings.HasPrefix(m, p) {
-			m = m[len(p):]
-			break
-		}
-	}
-	return bedrockVersionSuffix.ReplaceAllString(m, "")
-}

proxy/internal/llm/fixtures/pricing.yaml

@@ -1,59 +0,0 @@
-# Realistic-pricing starter for llm_observability. Drop this into the
-# directory you point the proxy at via --plugin-data-dir, then reference it
-# from the target's plugin config:
-#
-#   plugins:
-#     - id: llm_observability
-#       enabled: true
-#       params:
-#         pricing_path: pricing.yaml
-#
-# Values are USD per 1_000 tokens. Public list prices drift; treat this as a
-# starting point and keep your production copy current.
-
-openai:
-  # GPT-5 family
-  gpt-5:
-    input_per_1k: 0.00125
-    output_per_1k: 0.01
-  gpt-5-mini:
-    input_per_1k: 0.00025
-    output_per_1k: 0.002
-  gpt-5-nano:
-    input_per_1k: 0.00005
-    output_per_1k: 0.0004
-  gpt-5.4:
-    input_per_1k: 0.00125
-    output_per_1k: 0.01
-  # GPT-4o family
-  gpt-4o:
-    input_per_1k: 0.0025
-    output_per_1k: 0.01
-  gpt-4o-mini:
-    input_per_1k: 0.00015
-    output_per_1k: 0.0006
-  # Embeddings
-  text-embedding-3-large:
-    input_per_1k: 0.00013
-    output_per_1k: 0
-  text-embedding-3-small:
-    input_per_1k: 0.00002
-    output_per_1k: 0
-
-anthropic:
-  # Claude 4.x family
-  claude-opus-4-7:
-    input_per_1k: 0.015
-    output_per_1k: 0.075
-  claude-sonnet-4-7:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-  claude-sonnet-4-6:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-  claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-  claude-haiku-4-5:
-    input_per_1k: 0.0008
-    output_per_1k: 0.004

proxy/internal/llm/model.go

@@ -0,0 +1,21 @@
+package llm
+
+import (
+	sharedllm "github.com/netbirdio/netbird/shared/llm"
+)
+
+// NormalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
+// prefix, and the version/throughput suffix from a Bedrock model id so it
+// matches the catalog/pricing key. Thin delegate to the shared implementation
+// (shared/llm), which management also uses at synthesis time so both sides of
+// the pricing / routing contract normalize identically.
+func NormalizeBedrockModel(modelID string) string {
+	return sharedllm.NormalizeBedrockModel(modelID)
+}
+
+// NormalizeVertexModel strips the "@version" suffix from a Vertex AI model id
+// so it matches the catalog/pricing key. Thin delegate to shared/llm, kept
+// beside NormalizeBedrockModel for the same contract reason.
+func NormalizeVertexModel(modelID string) string {
+	return sharedllm.NormalizeVertexModel(modelID)
+}

proxy/internal/llm/pricing/defaults_coverage_test.go

@@ -1,65 +0,0 @@
-package pricing
-
-import (
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// TestDefaultTable_FirstPartyModelCoverage guards the embedded defaults against
-// silent drift/gaps: every metered first-party model the management catalog
-// enumerates must resolve to a price, and a few rates that previously drifted
-// are pinned to their LiteLLM-validated values. Keep this list in step with the
-// catalog (management/server/agentnetwork/catalog) when adding models.
-func TestDefaultTable_FirstPartyModelCoverage(t *testing.T) {
-	tbl := DefaultTable()
-	require.NotNil(t, tbl, "embedded default pricing table must load")
-
-	mustPrice := map[string][]string{
-		// openai parser covers openai_api, azure_openai_api, and mistral_api.
-		"openai": {
-			"gpt-5.5", "gpt-5.5-pro", "gpt-5.4", "gpt-5.4-mini", "gpt-5.4-nano",
-			"gpt-5.3-codex", "gpt-5.3-chat-latest", "o4-mini",
-			"gpt-4.1", "gpt-4.1-mini", "gpt-4.1-nano", "gpt-4o", "gpt-4o-mini",
-			"gpt-4-turbo", "gpt-3.5-turbo", "gpt-35-turbo",
-			"text-embedding-3-large", "text-embedding-3-small",
-			"mistral-large-latest", "mistral-medium-3-5", "codestral-2508",
-			"ministral-8b-latest", "mistral-embed",
-		},
-		"anthropic": {
-			"claude-fable-5", "claude-opus-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6",
-			"claude-opus-4-1", "claude-sonnet-4-6", "claude-sonnet-4-5", "claude-haiku-4-5",
-		},
-		// bedrock keys are the normalized ids the request parser emits.
-		"bedrock": {
-			"anthropic.claude-opus-5", "anthropic.claude-opus-4-8", "anthropic.claude-opus-4-7", "anthropic.claude-opus-4-6",
-			"anthropic.claude-opus-4-1", "anthropic.claude-sonnet-4-6", "anthropic.claude-sonnet-4-5",
-			"anthropic.claude-haiku-4-5", "meta.llama3-3-70b-instruct",
-			"amazon.nova-pro", "amazon.nova-lite", "amazon.nova-micro", "amazon.nova-2-lite",
-		},
-	}
-	for provider, models := range mustPrice {
-		for _, m := range models {
-			_, ok := tbl.Cost(provider, m, 1000, 1000, 0, 0)
-			assert.True(t, ok, "%s/%s must be priced in the embedded defaults", provider, m)
-		}
-	}
-
-	// Pin per-direction rates independently (input-only then output-only) so a
-	// swap or skew of input<->output that preserves the combined total is still
-	// caught — these are rates that previously drifted or are easy to mis-enter.
-	in, ok := tbl.Cost("openai", "gpt-5.4", 1000, 0, 0, 0)
-	require.True(t, ok)
-	assert.InDelta(t, 0.0025, in, 1e-9, "gpt-5.4 input = 0.0025 per 1k")
-	out, ok := tbl.Cost("openai", "gpt-5.4", 0, 1000, 0, 0)
-	require.True(t, ok)
-	assert.InDelta(t, 0.015, out, 1e-9, "gpt-5.4 output = 0.015 per 1k")
-
-	in, ok = tbl.Cost("bedrock", "anthropic.claude-sonnet-4-5", 1000, 0, 0, 0)
-	require.True(t, ok)
-	assert.InDelta(t, 0.003, in, 1e-9, "bedrock sonnet-4-5 input = 0.003 per 1k")
-	out, ok = tbl.Cost("bedrock", "anthropic.claude-sonnet-4-5", 0, 1000, 0, 0)
-	require.True(t, ok)
-	assert.InDelta(t, 0.015, out, 1e-9, "bedrock sonnet-4-5 output = 0.015 per 1k")
-}

proxy/internal/llm/pricing/pricing.go

@@ -1,102 +1,30 @@
-// Package pricing implements the embedded-default + override pricing table
-// shared by middleware that converts LLM token usage into a USD cost
-// estimate. The table is hot-reloadable from a basename under the proxy
-// data directory; missing override files keep the embedded defaults so
-// cost annotation works without operator action.
+// Package pricing implements the pricing table and cost formula the
+// cost_meter middleware uses to convert LLM token usage into a USD cost
+// estimate. The table's content arrives from the management server inside
+// cost_meter's middleware config (synthesized from the catalog plus the
+// operator's stored per-provider prices) — the proxy carries no embedded
+// price list. Price updates ride the ordinary mapping push: a chain
+// rebuild constructs a fresh table, so there is nothing to reload.
 package pricing
 
 import (
-	"bytes"
-	"context"
-	_ "embed"
-	"errors"
 	"fmt"
-	"io"
-	"io/fs"
 	"math"
-	"path/filepath"
-	"regexp"
-	"strings"
-	"sync"
-	"sync/atomic"
-	"time"
-
-	log "github.com/sirupsen/logrus"
-	"go.opentelemetry.io/otel/attribute"
-	"go.opentelemetry.io/otel/metric"
-	"gopkg.in/yaml.v3"
 )
 
-//go:embed defaults_pricing.yaml
-var defaultPricingYAML []byte
-
-var (
-	defaultTableOnce sync.Once
-	defaultTablePtr  *Table
-)
-
-// DefaultTable returns the pricing table embedded in the binary. The result
-// is parsed once and shared; callers must not mutate the returned value.
-// Cost annotation works without any operator action because every loader
-// starts with this table.
-func DefaultTable() *Table {
-	defaultTableOnce.Do(func() {
-		t, err := parsePricingBytes(defaultPricingYAML)
-		if err != nil {
-			panic(fmt.Sprintf("llmobs: embedded default pricing failed to parse: %v", err))
-		}
-		defaultTablePtr = t
-	})
-	return defaultTablePtr
-}
-
-// mergeOver returns a new Table containing every entry from base, with any
-// matching entry from overlay replacing the base value. Either argument may
-// be nil. Result is a fresh allocation so callers can mutate / Store safely.
-func mergeOver(base, overlay *Table) *Table {
-	if overlay == nil || len(overlay.entries) == 0 {
-		return base
-	}
-	if base == nil || len(base.entries) == 0 {
-		return overlay
-	}
-	out := make(map[string]map[string]Entry, len(base.entries))
-	for provider, models := range base.entries {
-		inner := make(map[string]Entry, len(models))
-		for model, e := range models {
-			inner[model] = e
-		}
-		out[provider] = inner
-	}
-	for provider, models := range overlay.entries {
-		inner, ok := out[provider]
-		if !ok {
-			inner = make(map[string]Entry, len(models))
-			out[provider] = inner
-		}
-		for model, e := range models {
-			inner[model] = e
-		}
-	}
-	return &Table{entries: out}
-}
-
 // Entry is a single model's input and output pricing, expressed in USD per
 // 1000 tokens.
 //
 // CachedInputPer1K applies to OpenAI's cached prompt tokens, which are a
 // subset of input_tokens — when set, the cached portion is billed at this
 // rate and the non-cached remainder at InputPer1K. Zero means "no discount
-// configured", and cached tokens are billed at InputPer1K (matches current
-// behaviour where cached counts weren't extracted at all).
+// configured", and cached tokens are billed at InputPer1K.
 //
 // CacheReadPer1K and CacheCreationPer1K apply to Anthropic's two prompt-
 // cache fields, which are additive to input_tokens: cache_read is the
 // cheaper read-from-cache rate, cache_creation is the more expensive
 // write-to-cache rate. Zero means "no rate configured" and the
-// corresponding token bucket is billed at InputPer1K. This is more
-// accurate than today's behaviour, where Anthropic's cache tokens are
-// ignored and not charged at all.
+// corresponding token bucket is billed at InputPer1K.
 type Entry struct {
 	InputPer1K         float64
 	OutputPer1K        float64
@@ -105,33 +33,102 @@ type Entry struct {
 	CacheCreationPer1K float64
 }
 
-// Table is a provider-to-model pricing lookup. Instances are immutable once
-// built and are swapped atomically by Loader.
+// EntryJSON is the wire shape of a pricing entry inside cost_meter's
+// middleware config. Field names are the management→proxy contract; the
+// management synthesizer marshals the same names (its pricing.Entry).
+type EntryJSON struct {
+	InputPer1K         float64 `json:"input_per_1k"`
+	OutputPer1K        float64 `json:"output_per_1k"`
+	CachedInputPer1K   float64 `json:"cached_input_per_1k"`
+	CacheReadPer1K     float64 `json:"cache_read_per_1k"`
+	CacheCreationPer1K float64 `json:"cache_creation_per_1k"`
+}
+
+// Table is a provider-surface-to-model pricing lookup. Instances are
+// immutable once built; a mapping update builds a whole new middleware
+// instance (and with it a new table) rather than mutating this one.
 type Table struct {
 	entries map[string]map[string]Entry
 }
 
+// NewEntries validates and converts a wire-shape map (surface-or-record ->
+// model -> rates) into the internal representation. Every rate must be a
+// finite, non-negative USD amount; a violation is returned as an error so
+// a corrupt config fails the chain build loudly instead of mispricing.
+// Management validates the same constraints at its API boundary, so this
+// is defense-in-depth. Nil input yields an empty (never-matching) map.
+func NewEntries(raw map[string]map[string]EntryJSON) (map[string]map[string]Entry, error) {
+	out := make(map[string]map[string]Entry, len(raw))
+	for outer, models := range raw {
+		inner := make(map[string]Entry, len(models))
+		for model, e := range models {
+			for field, v := range map[string]float64{
+				"input_per_1k":          e.InputPer1K,
+				"output_per_1k":         e.OutputPer1K,
+				"cached_input_per_1k":   e.CachedInputPer1K,
+				"cache_read_per_1k":     e.CacheReadPer1K,
+				"cache_creation_per_1k": e.CacheCreationPer1K,
+			} {
+				if v < 0 || math.IsNaN(v) || math.IsInf(v, 0) {
+					return nil, fmt.Errorf("pricing %s/%s: %s must be a finite, non-negative rate, got %v", outer, model, field, v)
+				}
+			}
+			// EntryJSON and Entry are field-identical (tags aside), so a
+			// direct conversion carries all five rates.
+			inner[model] = Entry(e)
+		}
+		out[outer] = inner
+	}
+	return out, nil
+}
+
+// NewTable builds an immutable Table from the wire-shape defaults map.
+// See NewEntries for validation semantics.
+func NewTable(raw map[string]map[string]EntryJSON) (*Table, error) {
+	entries, err := NewEntries(raw)
+	if err != nil {
+		return nil, err
+	}
+	return &Table{entries: entries}, nil
+}
+
+// Lookup returns the entry for the given provider surface and model.
+func (t *Table) Lookup(provider, model string) (Entry, bool) {
+	if t == nil {
+		return Entry{}, false
+	}
+	byModel, ok := t.entries[provider]
+	if !ok {
+		return Entry{}, false
+	}
+	e, ok := byModel[model]
+	return e, ok
+}
+
+// Has reports whether the provider/model pair is present in the table.
+func (t *Table) Has(provider, model string) bool {
+	_, ok := t.Lookup(provider, model)
+	return ok
+}
+
 // Cost returns the estimated USD cost for the given token counts. ok is
 // false when the provider or model is not present in the table; the caller
 // can still emit token metrics with a model=unknown label.
-//
-// Provider-shape semantics for cached / cache-creation counts:
-//
-//   - OpenAI: cachedInput is a SUBSET of inTokens. The cached portion is
-//     billed at CachedInputPer1K (or InputPer1K when no override), and the
-//     non-cached remainder of inTokens at InputPer1K. cacheCreation is
-//     ignored (OpenAI has no analogue).
-//   - Anthropic: cachedInput (cache_read) and cacheCreation are ADDITIVE to
-//     inTokens. The three buckets are billed at CacheReadPer1K,
-//     CacheCreationPer1K, and InputPer1K respectively, each falling back
-//     to InputPer1K when the corresponding rate is zero.
-//   - Other providers: cached and cacheCreation are ignored; cost is
-//     inTokens*InputPer1K + outTokens*OutputPer1K.
 func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (float64, bool) {
 	c, ok := t.Costs(provider, model, inTokens, outTokens, cachedInput, cacheCreation)
 	return c.TotalUSD, ok
 }
 
+// Costs returns the estimated USD cost split for the given token counts.
+// The provider surface selects the cache formula; see EntryCosts.
+func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (Costs, bool) {
+	entry, ok := t.Lookup(provider, model)
+	if !ok {
+		return Costs{}, false
+	}
+	return EntryCosts(entry, provider, inTokens, outTokens, cachedInput, cacheCreation), true
+}
+
 // Costs is a per-request cost split. The four per-bucket fields are the base
 // of the breakdown — one per token bucket the provider bills separately — and
 // the two aggregates are derived from them:
@@ -165,9 +162,25 @@ func newCosts(input, cachedInput, cacheCreation, output float64) Costs {
 	}
 }
 
-// Costs returns the estimated USD cost split for the given token counts, with
-// the same semantics as Cost.
-func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (Costs, bool) {
+// EntryCosts computes the USD cost split for the given entry and token
+// counts. The surface (the llm.provider value the request parser stamps)
+// selects the cache formula; the entry may come from the surface-keyed
+// defaults table or from a per-provider-record override — the math is
+// identical either way.
+//
+// Provider-shape semantics for cached / cache-creation counts:
+//
+//   - "openai": cachedInput is a SUBSET of inTokens. The cached portion is
+//     billed at CachedInputPer1K (or InputPer1K when no override), and the
+//     non-cached remainder of inTokens at InputPer1K. cacheCreation is
+//     ignored (OpenAI has no analogue).
+//   - "anthropic", "bedrock": cachedInput (cache_read) and cacheCreation are
+//     ADDITIVE to inTokens. The three buckets are billed at CacheReadPer1K,
+//     CacheCreationPer1K, and InputPer1K respectively, each falling back
+//     to InputPer1K when the corresponding rate is zero.
+//   - Other surfaces: cached and cacheCreation are ignored; cost is
+//     inTokens*InputPer1K + outTokens*OutputPer1K.
+func EntryCosts(entry Entry, surface string, inTokens, outTokens, cachedInput, cacheCreation int64) Costs {
 	// Clamp negatives to zero before any pricing math so a malformed
 	// upstream count can never produce a negative cost.
 	if inTokens < 0 {
@@ -182,19 +195,8 @@ func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput,
 	if cacheCreation < 0 {
 		cacheCreation = 0
 	}
-	if t == nil {
-		return Costs{}, false
-	}
-	byModel, ok := t.entries[provider]
-	if !ok {
-		return Costs{}, false
-	}
-	entry, ok := byModel[model]
-	if !ok {
-		return Costs{}, false
-	}
 	output := (float64(outTokens) / 1000.0) * entry.OutputPer1K
-	switch provider {
+	switch surface {
 	case "openai":
 		// cachedInput is a subset of inTokens; clamp so a malformed
 		// upstream (cached > total) can't produce a negative remainder.
@@ -208,7 +210,7 @@ func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput,
 		}
 		nonCached := float64(inTokens-clamped) / 1000.0 * entry.InputPer1K
 		cached := float64(clamped) / 1000.0 * cachedRate
-		return newCosts(nonCached, cached, 0, output), true
+		return newCosts(nonCached, cached, 0, output)
 	case "anthropic", "bedrock":
 		// Bedrock-Anthropic returns the same additive cache buckets as
 		// first-party Anthropic; non-Anthropic Bedrock models simply report
@@ -224,266 +226,9 @@ func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput,
 		input := float64(inTokens) / 1000.0 * entry.InputPer1K
 		read := float64(cachedInput) / 1000.0 * readRate
 		create := float64(cacheCreation) / 1000.0 * createRate
-		return newCosts(input, read, create, output), true
+		return newCosts(input, read, create, output)
 	default:
 		input := float64(inTokens) / 1000.0 * entry.InputPer1K
-		return newCosts(input, 0, 0, output), true
+		return newCosts(input, 0, 0, output)
 	}
 }
-
-// Has reports whether the provider/model pair is present in the table.
-func (t *Table) Has(provider, model string) bool {
-	if t == nil {
-		return false
-	}
-	byModel, ok := t.entries[provider]
-	if !ok {
-		return false
-	}
-	_, ok = byModel[model]
-	return ok
-}
-
-// pricingFile mirrors the on-disk YAML schema. Keys are provider names; the
-// nested map keys are model names.
-type pricingFile map[string]map[string]struct {
-	InputPer1K         float64 `yaml:"input_per_1k"`
-	OutputPer1K        float64 `yaml:"output_per_1k"`
-	CachedInputPer1K   float64 `yaml:"cached_input_per_1k"`
-	CacheReadPer1K     float64 `yaml:"cache_read_per_1k"`
-	CacheCreationPer1K float64 `yaml:"cache_creation_per_1k"`
-}
-
-const (
-	// ReloadInterval is the mtime-poll cadence for the background reloader.
-	ReloadInterval = 30 * time.Second
-
-	// errorBackoff bounds how often the loader logs a repeated parse error.
-	errorBackoff = 5 * time.Minute
-)
-
-var basenameRegex = regexp.MustCompile(`^[a-zA-Z0-9._-]+$`)
-
-// Loader is a confined, hot-reloadable pricing table reader. Construction
-// must succeed against the target file; subsequent reload failures keep the
-// previously-loaded table so callers never observe a blank price list.
-type Loader struct {
-	baseDir  string
-	fullPath string
-	pluginID string
-	table    atomic.Pointer[Table]
-	mtime    atomic.Int64
-	failures metric.Int64Counter
-	interval time.Duration
-}
-
-// NewLoader returns a pricing loader that overlays an optional file-based
-// table on top of the embedded defaults. Missing override file, baseDir, or
-// relPath is not an error: the loader keeps the embedded defaults so cost
-// metadata is still emitted for known models.
-//
-// Errors:
-//   - bad basename, traversal segment, or absolute relPath are rejected so a
-//     misconfigured target surfaces immediately.
-//   - permission errors and YAML parse errors keep the defaults but log a
-//     warning; cost annotation does not silently break.
-//
-// failures is optional; pass nil in tests that do not care about
-// reload-failure telemetry.
-func NewLoader(baseDir, relPath, pluginID string, failures metric.Int64Counter) (*Loader, error) {
-	defaults := DefaultTable()
-	l := &Loader{
-		baseDir:  baseDir,
-		pluginID: pluginID,
-		failures: failures,
-	}
-	l.table.Store(defaults)
-
-	if strings.TrimSpace(baseDir) == "" || strings.TrimSpace(relPath) == "" {
-		return l, nil
-	}
-
-	full, err := resolveMiddlewareDataPath(baseDir, relPath)
-	if err != nil {
-		return nil, err
-	}
-	l.fullPath = full
-
-	overlay, mtime, err := loadPricing(full)
-	if err != nil {
-		if errors.Is(err, fs.ErrNotExist) {
-			// Override file is optional. Defaults already stored.
-			return l, nil
-		}
-		// Symlink rejection, oversize file, parse failure, permission errors
-		// — surface so a misconfigured operator sees the problem instead of
-		// silently running with stale defaults.
-		return nil, fmt.Errorf("load pricing %s: %w", full, err)
-	}
-	l.table.Store(mergeOver(defaults, overlay))
-	l.mtime.Store(mtime.UnixNano())
-	return l, nil
-}
-
-// Get returns the current pricing table. The returned pointer is immutable;
-// callers must not mutate its contents.
-func (l *Loader) Get() *Table {
-	if l == nil {
-		return nil
-	}
-	return l.table.Load()
-}
-
-// WatchesFile reports whether this loader is bound to an override file on
-// disk. False for defaults-only loaders (no operator override given).
-// Callers use this to decide whether to spawn the mtime-poll goroutine.
-func (l *Loader) WatchesFile() bool {
-	if l == nil {
-		return false
-	}
-	return l.fullPath != ""
-}
-
-// SetReloadInterval overrides the mtime-poll cadence used by Reload. Calls
-// after Reload has started have no effect on the running loop. Intended for
-// tests; production code uses the default ReloadInterval.
-func (l *Loader) SetReloadInterval(d time.Duration) {
-	if l == nil || d <= 0 {
-		return
-	}
-	l.interval = d
-}
-
-// Reload runs a polling loop that checks the pricing file mtime every
-// ReloadInterval (or the value passed to SetReloadInterval). Returns when
-// ctx is cancelled.
-func (l *Loader) Reload(ctx context.Context) {
-	if l == nil {
-		return
-	}
-	interval := l.interval
-	if interval <= 0 {
-		interval = ReloadInterval
-	}
-	t := time.NewTicker(interval)
-	defer t.Stop()
-
-	var lastErrAt time.Time
-	for {
-		select {
-		case <-ctx.Done():
-			return
-		case <-t.C:
-			if err := l.reload(); err != nil {
-				if l.failures != nil {
-					l.failures.Add(ctx, 1, metric.WithAttributes(
-						attribute.String("plugin", l.pluginID),
-					))
-				}
-				now := time.Now()
-				if now.Sub(lastErrAt) >= errorBackoff {
-					log.Warnf("llmobs: pricing reload failed for %s: %v", l.fullPath, err)
-					lastErrAt = now
-				}
-			}
-		}
-	}
-}
-
-// reload performs a single-shot mtime check and reload. The reloaded
-// override file is merged on top of the embedded defaults; missing override
-// (e.g. operator deleted the file) is not an error and reverts to defaults.
-func (l *Loader) reload() error {
-	if l.fullPath == "" {
-		// Defaults-only loader; nothing on disk to reload.
-		return nil
-	}
-	mtime, err := statMtime(l.fullPath)
-	if err != nil {
-		if errors.Is(err, fs.ErrNotExist) {
-			// File was removed since startup. Drop back to defaults and
-			// reset mtime so a future re-creation triggers a reload.
-			l.table.Store(DefaultTable())
-			l.mtime.Store(0)
-			return nil
-		}
-		return err
-	}
-	if mtime.UnixNano() == l.mtime.Load() {
-		return nil
-	}
-
-	overlay, newMtime, err := loadPricing(l.fullPath)
-	if err != nil {
-		return err
-	}
-	l.table.Store(mergeOver(DefaultTable(), overlay))
-	l.mtime.Store(newMtime.UnixNano())
-	return nil
-}
-
-// resolveMiddlewareDataPath validates relPath is a safe basename and resolves
-// it under baseDir. An additional cleaned-prefix check guards against
-// CVE-style edge cases where Join is used with trailing path segments.
-func resolveMiddlewareDataPath(baseDir, relPath string) (string, error) {
-	if strings.TrimSpace(baseDir) == "" {
-		return "", errors.New("middleware-data-dir is not configured")
-	}
-	if relPath == "" {
-		return "", errors.New("pricing path is empty")
-	}
-	if !basenameRegex.MatchString(relPath) {
-		return "", fmt.Errorf("pricing path %q is not a safe basename", relPath)
-	}
-	if filepath.IsAbs(relPath) {
-		return "", fmt.Errorf("pricing path %q must be a basename, not absolute", relPath)
-	}
-
-	cleanBase, err := filepath.Abs(filepath.Clean(baseDir))
-	if err != nil {
-		return "", fmt.Errorf("resolve middleware-data-dir: %w", err)
-	}
-	full := filepath.Join(cleanBase, relPath)
-	cleanedFull := filepath.Clean(full)
-	if !strings.HasPrefix(cleanedFull, cleanBase+string(filepath.Separator)) && cleanedFull != cleanBase {
-		return "", fmt.Errorf("pricing path %q escapes middleware-data-dir", relPath)
-	}
-	return cleanedFull, nil
-}
-
-func parsePricingBytes(data []byte) (*Table, error) {
-	dec := yaml.NewDecoder(bytes.NewReader(data))
-	dec.KnownFields(true)
-
-	var raw pricingFile
-	if err := dec.Decode(&raw); err != nil && !errors.Is(err, io.EOF) {
-		return nil, fmt.Errorf("decode pricing yaml: %w", err)
-	}
-
-	out := make(map[string]map[string]Entry, len(raw))
-	for provider, models := range raw {
-		inner := make(map[string]Entry, len(models))
-		for model, entry := range models {
-			for field, v := range map[string]float64{
-				"input_per_1k":          entry.InputPer1K,
-				"output_per_1k":         entry.OutputPer1K,
-				"cached_input_per_1k":   entry.CachedInputPer1K,
-				"cache_read_per_1k":     entry.CacheReadPer1K,
-				"cache_creation_per_1k": entry.CacheCreationPer1K,
-			} {
-				if v < 0 || math.IsNaN(v) || math.IsInf(v, 0) {
-					return nil, fmt.Errorf("pricing %s/%s: %s must be a finite, non-negative rate, got %v", provider, model, field, v)
-				}
-			}
-			inner[model] = Entry{
-				InputPer1K:         entry.InputPer1K,
-				OutputPer1K:        entry.OutputPer1K,
-				CachedInputPer1K:   entry.CachedInputPer1K,
-				CacheReadPer1K:     entry.CacheReadPer1K,
-				CacheCreationPer1K: entry.CacheCreationPer1K,
-			}
-		}
-		out[provider] = inner
-	}
-	return &Table{entries: out}, nil
-}

proxy/internal/llm/pricing/pricing_other.go

@@ -1,20 +0,0 @@
-//go:build !unix
-
-package pricing
-
-import (
-	"fmt"
-	"time"
-)
-
-// loadPricing is unavailable on non-Unix platforms because O_NOFOLLOW and
-// fstat-from-FD are required to honour the spec's symlink-safety rules. The
-// proxy is only deployed on Linux today; a Windows port would need an
-// equivalent path-as-handle implementation.
-func loadPricing(path string) (*Table, time.Time, error) {
-	return nil, time.Time{}, fmt.Errorf("llmobs pricing loader is not supported on this platform: %s", path)
-}
-
-func statMtime(path string) (time.Time, error) {
-	return time.Time{}, fmt.Errorf("llmobs pricing loader is not supported on this platform: %s", path)
-}

proxy/internal/llm/pricing/pricing_test.go

@@ -1,47 +1,13 @@
-//go:build unix
-
 package pricing
 
 import (
-	"context"
-	"os"
-	"path/filepath"
+	"math"
 	"testing"
-	"time"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
 
-func copyFixture(t *testing.T, src, dst string) {
-	t.Helper()
-	data, err := os.ReadFile(src)
-	require.NoError(t, err, "read source fixture")
-	require.NoError(t, os.WriteFile(dst, data, 0o600), "write target fixture")
-}
-
-func TestNewLoader_HappyPath(t *testing.T) {
-	base := t.TempDir()
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
-
-	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.NoError(t, err, "NewLoader must succeed with a valid fixture")
-	table := l.Get()
-	require.NotNil(t, table, "table populated after load")
-
-	cost, ok := table.Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
-	require.True(t, ok, "known provider/model resolves")
-	assert.InDelta(t, 0.00075, cost, 1e-9, "cost = 0.00015 + 0.0006 per 1k tokens")
-
-	cost, ok = table.Cost("openai", "gpt-4o", 2000, 1000, 0, 0)
-	require.True(t, ok, "second known model resolves")
-	assert.InDelta(t, 0.015, cost, 1e-9, "cost for gpt-4o: 2*0.0025 + 1*0.01")
-
-	cost, ok = table.Cost("anthropic", "claude-sonnet-4-5", 1000, 1000, 0, 0)
-	require.True(t, ok, "anthropic model resolves")
-	assert.InDelta(t, 0.018, cost, 1e-9, "cost for claude-sonnet-4-5: 0.003 + 0.015")
-}
-
 // TestCost_OpenAICachedSubsetDiscount proves OpenAI's cached input
 // tokens are billed at the configured cached_input_per_1k rate while
 // the non-cached remainder of input_tokens is billed at the regular
@@ -65,11 +31,9 @@ func TestCost_OpenAICachedSubsetDiscount(t *testing.T) {
 		"cached subset must bill at the discount rate; non-cached remainder at regular rate")
 }
 
-// TestCost_OpenAICachedFallsBackToInputRate covers the operator
-// opt-in contract: when CachedInputPer1K is unset (zero), cached
-// tokens bill at the regular input rate. This matches today's
-// behaviour (cached counts weren't extracted at all so they
-// implicitly billed at the input rate via prompt_tokens).
+// TestCost_OpenAICachedFallsBackToInputRate covers the fallback
+// contract: when CachedInputPer1K is unset (zero), cached tokens bill
+// at the regular input rate.
 func TestCost_OpenAICachedFallsBackToInputRate(t *testing.T) {
 	tbl := &Table{entries: map[string]map[string]Entry{
 		"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01}},
@@ -78,7 +42,7 @@ func TestCost_OpenAICachedFallsBackToInputRate(t *testing.T) {
 	require.True(t, ok)
 	want := 0.0025 + (500.0/1000.0)*0.01
 	assert.InDelta(t, want, cost, 1e-12,
-		"absent cached_input_per_1k rate must fall back to input_per_1k — same as pre-feature behaviour")
+		"absent cached_input_per_1k rate must fall back to input_per_1k")
 }
 
 // TestCost_OpenAIClampsCachedToInputCount is the defensive guard
@@ -100,10 +64,7 @@ func TestCost_OpenAIClampsCachedToInputCount(t *testing.T) {
 // TestCost_AnthropicCacheReadAndCreationAreAdditive proves the
 // Anthropic shape: cache_read and cache_creation tokens are
 // ADDITIVE to input_tokens (not subset), each billed at its own
-// configured rate. The two rates pull in opposite directions —
-// cache_read is the cheaper read-from-cache rate (≈0.1× input),
-// cache_creation is the more expensive write-to-cache rate
-// (≈1.25× input).
+// configured rate.
 func TestCost_AnthropicCacheReadAndCreationAreAdditive(t *testing.T) {
 	tbl := &Table{entries: map[string]map[string]Entry{
 		"anthropic": {"claude-sonnet": {
@@ -125,11 +86,9 @@ func TestCost_AnthropicCacheReadAndCreationAreAdditive(t *testing.T) {
 		"each Anthropic input bucket must bill at its own configured rate")
 }
 
-// TestCost_AnthropicCacheRatesFallBackToInput covers the no-opt-in
+// TestCost_AnthropicCacheRatesFallBackToInput covers the no-rate
 // path: when neither CacheReadPer1K nor CacheCreationPer1K is set,
-// cache tokens bill at the regular input rate. This is more
-// accurate than today's behaviour (cache tokens ignored entirely)
-// without requiring operators to opt in via YAML.
+// cache tokens bill at the regular input rate.
 func TestCost_AnthropicCacheRatesFallBackToInput(t *testing.T) {
 	tbl := &Table{entries: map[string]map[string]Entry{
 		"anthropic": {"claude-sonnet": {InputPer1K: 0.003, OutputPer1K: 0.015}},
@@ -139,259 +98,39 @@ func TestCost_AnthropicCacheRatesFallBackToInput(t *testing.T) {
 	// Without overrides: every input bucket at input_per_1k.
 	want := ((256.0+768.0+512.0)/1000.0)*0.003 + (200.0/1000.0)*0.015
 	assert.InDelta(t, want, cost, 1e-12,
-		"absent cache rates must fall back to input_per_1k — Anthropic cache tokens were ignored before this change, billing at input rate is more accurate as a default")
+		"absent cache rates must fall back to input_per_1k")
 }
 
-func TestNewLoader_UnknownModel(t *testing.T) {
-	base := t.TempDir()
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
+// TestEntryCosts_SurfaceSelectsFormula pins that the formula branches on
+// the SURFACE, not on which table the entry came from: the same entry
+// bills a subset carve-out on "openai", additive buckets on
+// "anthropic"/"bedrock", and ignores cache counts everywhere else. This
+// is what keeps per-provider-record entries (looked up by record id)
+// mathematically identical to defaults-table entries.
+func TestEntryCosts_SurfaceSelectsFormula(t *testing.T) {
+	e := Entry{InputPer1K: 0.002, OutputPer1K: 0.01, CachedInputPer1K: 0.001, CacheReadPer1K: 0.0002, CacheCreationPer1K: 0.0025}
 
-	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.NoError(t, err)
+	openai := EntryCosts(e, "openai", 1000, 0, 400, 300)
+	assert.InDelta(t, (600.0/1000.0)*0.002+(400.0/1000.0)*0.001, openai.TotalUSD, 1e-12,
+		"openai: cached is a subset, cacheCreation ignored")
 
-	_, ok := l.Get().Cost("openai", "fantasy-model", 10, 10, 0, 0)
-	assert.False(t, ok, "unknown model returns ok=false")
+	anthropic := EntryCosts(e, "anthropic", 1000, 0, 400, 300)
+	assert.InDelta(t, 0.002+(400.0/1000.0)*0.0002+(300.0/1000.0)*0.0025, anthropic.TotalUSD, 1e-12,
+		"anthropic: cache buckets are additive")
 
-	_, ok = l.Get().Cost("cohere", "anything", 10, 10, 0, 0)
-	assert.False(t, ok, "unknown provider returns ok=false")
+	bedrock := EntryCosts(e, "bedrock", 1000, 0, 400, 300)
+	assert.InDelta(t, anthropic.TotalUSD, bedrock.TotalUSD, 1e-12, "bedrock shares the anthropic formula")
+
+	other := EntryCosts(e, "gemini", 1000, 0, 400, 300)
+	assert.InDelta(t, 0.002, other.TotalUSD, 1e-12, "unknown surface: cache counts ignored")
 }
 
-func TestNewLoader_InvalidYAMLRejected(t *testing.T) {
-	base := t.TempDir()
-	require.NoError(t, os.WriteFile(filepath.Join(base, "pricing.yaml"), []byte("\t- this is not: valid: yaml: :["), 0o600))
-
-	_, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.Error(t, err, "invalid YAML must surface as construction error")
-}
-
-func TestLoader_ReloadKeepsPreviousOnParseError(t *testing.T) {
-	base := t.TempDir()
-	target := filepath.Join(base, "pricing.yaml")
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
-
-	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.NoError(t, err)
-	require.NotNil(t, l.Get(), "initial table populated")
-
-	// Overwrite with content that violates the strict schema (extra field)
-	// plus a bumped mtime to trigger reload.
-	require.NoError(t, os.WriteFile(target, []byte("openai:\n  gpt-4o:\n    input_per_1k: 1.0\n    output_per_1k: 2.0\n    bogus_field: nope\n"), 0o600))
-	future := time.Now().Add(time.Hour)
-	require.NoError(t, os.Chtimes(target, future, future))
-
-	err = l.reload()
-	require.Error(t, err, "parse error surfaced by reload()")
-
-	cost, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
-	require.True(t, ok, "previous table still available after parse failure")
-	assert.InDelta(t, 0.00075, cost, 1e-9, "previous cost preserved")
-}
-
-func TestLoader_ReloadNoChangeIsNoOp(t *testing.T) {
-	base := t.TempDir()
-	target := filepath.Join(base, "pricing.yaml")
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
-
-	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.NoError(t, err)
-	ptrBefore := l.Get()
-
-	require.NoError(t, l.reload(), "no-change reload must not error")
-	ptrAfter := l.Get()
-	assert.Same(t, ptrBefore, ptrAfter, "table pointer unchanged when mtime unchanged")
-}
-
-func TestLoader_ReloadDetectsChange(t *testing.T) {
-	base := t.TempDir()
-	target := filepath.Join(base, "pricing.yaml")
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
-
-	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.NoError(t, err)
-
-	updated := []byte("openai:\n  gpt-4o-mini:\n    input_per_1k: 1.00\n    output_per_1k: 2.00\n")
-	require.NoError(t, os.WriteFile(target, updated, 0o600))
-	future := time.Now().Add(time.Hour)
-	require.NoError(t, os.Chtimes(target, future, future))
-
-	require.NoError(t, l.reload(), "reload must succeed on valid new content")
-
-	cost, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
-	require.True(t, ok, "updated model still present")
-	assert.InDelta(t, 3.0, cost, 0.0001, "new prices are applied: 1 + 2 per 1k")
-}
-
-// TestLoader_ReloadGoroutinePicksUpChanges proves the background goroutine
-// started via Reload actually swaps the pricing table when the file changes
-// on disk. Without that goroutine running, pricing edits would never reach
-// requests until a proxy restart.
-func TestLoader_ReloadGoroutinePicksUpChanges(t *testing.T) {
-	base := t.TempDir()
-	target := filepath.Join(base, "pricing.yaml")
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
-
-	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.NoError(t, err)
-	l.SetReloadInterval(20 * time.Millisecond)
-
-	ctx, cancel := context.WithCancel(context.Background())
-	defer cancel()
-	done := make(chan struct{})
-	go func() {
-		l.Reload(ctx)
-		close(done)
-	}()
-
-	// Before any rewrite, the loader holds the fixture's prices.
-	costBefore, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
-	require.True(t, ok, "fixture model must resolve initially")
-	assert.InDelta(t, 0.00075, costBefore, 1e-9, "fixture prices apply before rewrite")
-
-	updated := []byte("openai:\n  gpt-4o-mini:\n    input_per_1k: 1.00\n    output_per_1k: 2.00\n")
-	require.NoError(t, os.WriteFile(target, updated, 0o600))
-	future := time.Now().Add(time.Hour)
-	require.NoError(t, os.Chtimes(target, future, future))
-
-	deadline := time.Now().Add(2 * time.Second)
-	for {
-		cost, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
-		if ok && cost > 2.5 {
-			break
-		}
-		if time.Now().After(deadline) {
-			t.Fatalf("background reloader did not pick up rewrite within deadline")
-		}
-		time.Sleep(10 * time.Millisecond)
-	}
-
-	cancel()
-	select {
-	case <-done:
-	case <-time.After(2 * time.Second):
-		t.Fatal("Reload loop did not exit after cancel")
-	}
-}
-
-func TestLoader_ReloadBackgroundLoopCancellation(t *testing.T) {
-	base := t.TempDir()
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
-	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.NoError(t, err)
-
-	ctx, cancel := context.WithCancel(context.Background())
-	done := make(chan struct{})
-	go func() {
-		l.Reload(ctx)
-		close(done)
-	}()
-	cancel()
-
-	select {
-	case <-done:
-	case <-time.After(2 * time.Second):
-		t.Fatal("Reload loop did not exit on context cancel")
-	}
-}
-
-func TestNewLoader_PathValidation(t *testing.T) {
-	base := t.TempDir()
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
-
-	cases := []struct {
-		name    string
-		relPath string
-	}{
-		{"traversal", "../../etc/passwd"},
-		{"absolute", "/etc/passwd"},
-		{"slash in basename", "sub/pricing.yaml"},
-		{"control chars", "pricing\x00.yaml"},
-	}
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			_, err := NewLoader(base, tc.relPath, "llm_observability", nil)
-			require.Error(t, err, "NewLoader must reject %q", tc.relPath)
-		})
-	}
-
-	// Empty relPath is no longer a validation error: the loader treats it
-	// as "no override file, defaults only" so cost metadata is still
-	// emitted for the embedded models out of the box.
-	t.Run("empty falls back to defaults", func(t *testing.T) {
-		l, err := NewLoader(base, "", "llm_observability", nil)
-		require.NoError(t, err, "empty relPath should yield a defaults-only loader")
-		require.NotNil(t, l, "loader must be returned")
-		require.False(t, l.WatchesFile(), "no file watching when no override is given")
-		_, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
-		assert.True(t, ok, "embedded defaults should still resolve gpt-4o-mini")
-	})
-}
-
-// TestNewLoader_PathValidation_Extended covers the remaining attack shapes
-// called out in C2: dot references, embedded traversal segments, and a
-// newline in the basename. The basename regex must reject each one even
-// though filepath.Clean would otherwise collapse them.
-func TestNewLoader_PathValidation_Extended(t *testing.T) {
-	base := t.TempDir()
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
-
-	cases := []struct {
-		name    string
-		relPath string
-	}{
-		{"dot", "."},
-		{"dotdot", ".."},
-		{"relative traversal", "../pricing.yaml"},
-		{"embedded slash", "pri/cing.yaml"},
-		{"newline", "pricing\n.yaml"},
-	}
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			_, err := NewLoader(base, tc.relPath, "llm_observability", nil)
-			require.Error(t, err, "NewLoader must reject %q", tc.relPath)
-		})
-	}
-}
-
-// TestNewLoader_ValidBasenameLoads proves the allowlist is exclusive: a
-// basename containing only safe characters under baseDir loads. Without this
-// a regression that over-tightened the regex would silently break valid
-// deployments.
-func TestNewLoader_ValidBasenameLoads(t *testing.T) {
-	base := t.TempDir()
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing-v2_prod.yaml"))
-
-	l, err := NewLoader(base, "pricing-v2_prod.yaml", "llm_observability", nil)
-	require.NoError(t, err, "basename with _, -, . must load")
-	require.NotNil(t, l.Get(), "table populated")
-}
-
-// TestNewLoader_SymlinkOutsideBaseDirRejected constructs a symlink under
-// baseDir that points to a file outside it. O_NOFOLLOW must refuse to open
-// the symlink even though the symlink path itself is a valid basename under
-// baseDir.
-func TestNewLoader_SymlinkOutsideBaseDirRejected(t *testing.T) {
-	outside := t.TempDir()
-	target := filepath.Join(outside, "evil.yaml")
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
-
-	base := t.TempDir()
-	link := filepath.Join(base, "pricing.yaml")
-	require.NoError(t, os.Symlink(target, link), "symlink setup")
-
-	_, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.Error(t, err, "O_NOFOLLOW must reject symlink even when it points outside baseDir")
-}
-
-func TestNewLoader_SymlinkRejected(t *testing.T) {
-	base := t.TempDir()
-	concrete := filepath.Join(base, "real.yaml")
-	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), concrete)
-
-	link := filepath.Join(base, "pricing.yaml")
-	require.NoError(t, os.Symlink(concrete, link), "symlink setup")
-
-	_, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.Error(t, err, "O_NOFOLLOW must reject symlinked targets")
+// TestEntryCosts_ClampsNegativeTokens: malformed upstream counts must
+// never produce a negative cost.
+func TestEntryCosts_ClampsNegativeTokens(t *testing.T) {
+	e := Entry{InputPer1K: 0.002, OutputPer1K: 0.01}
+	c := EntryCosts(e, "openai", -50, -10, -5, -3)
+	assert.Zero(t, c.TotalUSD, "all-negative counts clamp to zero cost")
 }
 
 func TestTableCost_NilSafe(t *testing.T) {
@@ -402,31 +141,37 @@ func TestTableCost_NilSafe(t *testing.T) {
 	assert.False(t, t1.Has("x", "y"), "nil table has nothing")
 }
 
-func TestLoaderGet_NilSafe(t *testing.T) {
-	var l *Loader
-	assert.Nil(t, l.Get(), "nil loader returns nil table")
-}
-
-// TestNewLoader_RejectsOversizedFile_FixesM4 proves the loader bounds reads
-// at maxPricingBytes so a hostile file cannot exhaust process memory.
-func TestNewLoader_RejectsOversizedFile_FixesM4(t *testing.T) {
-	base := t.TempDir()
-	target := filepath.Join(base, "pricing.yaml")
-
-	// Build a YAML payload larger than the cap. We pad with valid YAML
-	// comments so a partial read would still fail the size check rather
-	// than the parser.
-	header := "openai:\n"
-	bigComment := make([]byte, maxPricingBytes+1024)
-	for i := range bigComment {
-		bigComment[i] = ' '
+func TestNewTable_ValidatesRates(t *testing.T) {
+	good := map[string]map[string]EntryJSON{
+		"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01, CachedInputPer1K: 0.00125}},
 	}
-	bigComment[0] = '#'
-	bigComment[len(bigComment)-1] = '\n'
-	payload := append([]byte(header), bigComment...)
-	require.NoError(t, os.WriteFile(target, payload, 0o600))
+	tbl, err := NewTable(good)
+	require.NoError(t, err)
+	cost, ok := tbl.Cost("openai", "gpt-4o", 1000, 1000, 0, 0)
+	require.True(t, ok, "entry survives the wire conversion")
+	assert.InDelta(t, 0.0125, cost, 1e-9)
 
-	_, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
-	require.Error(t, err, "oversized pricing file must be rejected")
-	assert.Contains(t, err.Error(), "exceeds", "rejection must reference the byte cap")
+	_, ok = tbl.Cost("openai", "unknown-model", 1, 1, 0, 0)
+	assert.False(t, ok, "unknown model misses")
+
+	for name, bad := range map[string]EntryJSON{
+		"negative input":  {InputPer1K: -1, OutputPer1K: 0.01},
+		"NaN output":      {InputPer1K: 0.01, OutputPer1K: math.NaN()},
+		"Inf cache read":  {InputPer1K: 0.01, OutputPer1K: 0.01, CacheReadPer1K: math.Inf(1)},
+		"negative cached": {InputPer1K: 0.01, OutputPer1K: 0.01, CachedInputPer1K: -0.001},
+	} {
+		_, err := NewTable(map[string]map[string]EntryJSON{"openai": {"m": bad}})
+		assert.Error(t, err, "case %q must be rejected so a corrupt config fails the chain build instead of mispricing", name)
+	}
+}
+
+func TestNewTable_NilAndEmpty(t *testing.T) {
+	tbl, err := NewTable(nil)
+	require.NoError(t, err, "nil map builds an empty (never-matching) table")
+	_, ok := tbl.Cost("openai", "gpt-4o", 1, 1, 0, 0)
+	assert.False(t, ok, "empty table prices nothing")
+
+	entries, err := NewEntries(nil)
+	require.NoError(t, err)
+	assert.Empty(t, entries, "nil in, empty (never-matching) map out for the per-record map")
 }

proxy/internal/llm/pricing/pricing_unix.go

@@ -1,68 +0,0 @@
-//go:build unix
-
-package pricing
-
-import (
-	"fmt"
-	"io"
-	"os"
-	"syscall"
-	"time"
-
-	log "github.com/sirupsen/logrus"
-)
-
-// maxPricingBytes caps the size of the pricing YAML on read so a hostile or
-// runaway file cannot exhaust process memory during reload. 1 MiB is several
-// orders of magnitude larger than any reasonable pricing table.
-const maxPricingBytes int64 = 1 << 20
-
-// loadPricing opens the file with O_NOFOLLOW, fstats the open descriptor,
-// and parses from that same descriptor. Never re-opens by path so a
-// mid-read rename or symlink swap cannot substitute content. Bytes are
-// capped at maxPricingBytes so the loader cannot be coerced into reading an
-// unbounded file.
-func loadPricing(path string) (*Table, time.Time, error) {
-	f, err := os.OpenFile(path, os.O_RDONLY|syscall.O_NOFOLLOW, 0)
-	if err != nil {
-		return nil, time.Time{}, fmt.Errorf("open %s: %w", path, err)
-	}
-	defer func() {
-		if cerr := f.Close(); cerr != nil {
-			log.Debugf("close pricing file %s: %v", path, cerr)
-		}
-	}()
-
-	info, err := f.Stat()
-	if err != nil {
-		return nil, time.Time{}, fmt.Errorf("fstat %s: %w", path, err)
-	}
-	if !info.Mode().IsRegular() {
-		return nil, time.Time{}, fmt.Errorf("pricing file %s is not a regular file", path)
-	}
-
-	data, err := io.ReadAll(io.LimitReader(f, maxPricingBytes+1))
-	if err != nil {
-		return nil, time.Time{}, fmt.Errorf("read %s: %w", path, err)
-	}
-	if int64(len(data)) > maxPricingBytes {
-		return nil, time.Time{}, fmt.Errorf("pricing file %s exceeds %d bytes", path, maxPricingBytes)
-	}
-
-	table, err := parsePricingBytes(data)
-	if err != nil {
-		return nil, time.Time{}, err
-	}
-	return table, info.ModTime(), nil
-}
-
-// statMtime returns the mtime of the file at path. It uses lstat semantics
-// via os.Lstat so a symlink swap is detected even though O_NOFOLLOW will
-// later reject the open.
-func statMtime(path string) (time.Time, error) {
-	info, err := os.Lstat(path)
-	if err != nil {
-		return time.Time{}, fmt.Errorf("lstat %s: %w", path, err)
-	}
-	return info.ModTime(), nil
-}

proxy/internal/middleware/builtin/builtin.go

@@ -36,15 +36,13 @@ var defaultRegistry = middleware.NewRegistry()
 
 // FactoryContext is the per-process bag that concrete factories may
 // consult during construction. It carries the proxy-lifetime context,
-// the data directory used for static config files (pricing tables,
-// allowlists), the OTel meter, and the proxy logger.
+// the OTel meter, and the proxy logger.
 //
 // Configure must be called once at boot before any chain build calls
 // Resolve. Calling it twice overwrites the prior value; tests may rely
 // on this to reset state.
 type FactoryContext struct {
 	Context    context.Context
-	DataDir    string
 	Meter      metric.Meter
 	Logger     *log.Logger
 	MgmtClient MgmtClient
@@ -58,12 +56,11 @@ var (
 // Configure stores the per-process FactoryContext. Concrete factories
 // reach for it via Context(). mgmt may be nil on tests / standalone
 // builds with no management server; consumers must guard.
-func Configure(ctx context.Context, dataDir string, meter metric.Meter, logger *log.Logger, mgmt MgmtClient) {
+func Configure(ctx context.Context, meter metric.Meter, logger *log.Logger, mgmt MgmtClient) {
 	ctxMu.Lock()
 	defer ctxMu.Unlock()
 	ctxStore = FactoryContext{
 		Context:    ctx,
-		DataDir:    dataDir,
 		Meter:      meter,
 		Logger:     logger,
 		MgmtClient: mgmt,

proxy/internal/middleware/builtin/cost_calculation_matrix_test.go

@@ -12,6 +12,7 @@ import (
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 
+	mgmtpricing "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/cost_meter"
@@ -19,17 +20,20 @@ import (
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_response_parser"
 )
 
-// Drives the real pipeline (llm_request_parser → llm_response_parser → cost_meter) on the embedded default pricing
-// table and asserts exact USD amounts hardcoded from the vendors' published prices, including the cache split.
+// Drives the real pipeline (llm_request_parser → llm_response_parser → cost_meter) on the REAL default pricing
+// table management ships (mgmtpricing.DefaultTable, catalog-derived) and asserts exact USD amounts hardcoded from
+// the vendors' published prices, including the cache split. This is the cross-stack pricing contract test: the
+// management-side Entry JSON must decode into the proxy-side table and produce these exact costs.
 func TestCostCalculation_ProviderMatrix(t *testing.T) {
-	// Empty data dir → embedded defaults, like a proxy with no pricing override.
-	builtin.Configure(context.Background(), t.TempDir(), nil, nil, nil)
+	builtin.Configure(context.Background(), nil, nil, nil)
 
 	reqMW, err := llm_request_parser.Factory{}.New(nil)
 	require.NoError(t, err, "build llm_request_parser")
 	respMW, err := llm_response_parser.Factory{}.New(nil)
 	require.NoError(t, err, "build llm_response_parser")
-	costMW, err := cost_meter.Factory{}.New(nil)
+	costCfgJSON, err := json.Marshal(map[string]any{"pricing": map[string]any{"defaults": mgmtpricing.DefaultTable()}})
+	require.NoError(t, err, "marshal management default table into cost_meter config")
+	costMW, err := cost_meter.Factory{}.New(costCfgJSON)
 	require.NoError(t, err, "build cost_meter")
 	t.Cleanup(func() { _ = costMW.Close() })
 

proxy/internal/middleware/builtin/cost_meter/factory.go

@@ -2,7 +2,6 @@ package cost_meter
 
 import (
 	"bytes"
-	"context"
 	"encoding/json"
 	"fmt"
 
@@ -11,16 +10,27 @@ import (
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 )
 
-// defaultPricingFilename is the basename probed inside the proxy data
-// directory when no override is configured.
-const defaultPricingFilename = "pricing.yaml"
-
-// Config is the on-wire configuration for the middleware.
+// Config is the on-wire configuration for the middleware, synthesized by
+// management (buildCostMeterConfigJSON). The proxy has no embedded price
+// list: this payload is the only pricing source, and updates arrive as
+// ordinary mapping pushes that rebuild the chain (and with it this
+// middleware instance) — no per-request fetches, no reload loops.
 type Config struct {
-	// PricingPath optionally overrides the basename of the pricing
-	// file probed inside the proxy data directory. When empty the
-	// loader falls back to "pricing.yaml".
-	PricingPath string `json:"pricing_path"`
+	Pricing *PricingConfig `json:"pricing"`
+}
+
+// PricingConfig carries the full pricing table:
+//   - Defaults: parser surface ("openai"/"anthropic"/"bedrock") ->
+//     normalized model id -> rates, matched against llm.provider +
+//     llm.model.
+//   - Providers: provider record id -> normalized model id -> rates,
+//     matched against the llm.resolved_provider_id metadata llm_router
+//     stamps. Entries arrive fully materialized (management folds default
+//     cache rates in at synth time), so lookup order is simply
+//     per-record first, defaults second.
+type PricingConfig struct {
+	Defaults  map[string]map[string]pricing.EntryJSON `json:"defaults"`
+	Providers map[string]map[string]pricing.EntryJSON `json:"providers"`
 }
 
 // Factory builds cost_meter instances from raw config bytes.
@@ -29,45 +39,45 @@ type Factory struct{}
 // ID returns the registry identifier.
 func (Factory) ID() string { return ID }
 
-// New constructs a middleware instance. Empty, null, and {} configs
-// are accepted; non-empty rawConfig that fails to unmarshal is
-// rejected so misconfigurations surface at chain build time. The
-// pricing loader is built once per instance and reused across
-// invocations.
+// New constructs a middleware instance. Empty, null, and {} configs are
+// accepted for backward compatibility with a management server that
+// predates config-delivered pricing — the instance then skips every cost
+// computation (unknown_model) and a warning is logged once at build time.
+// Non-empty rawConfig that fails to unmarshal, or a table carrying a
+// non-finite / negative rate, is rejected so misconfigurations surface at
+// chain build time.
 func (Factory) New(rawConfig []byte) (middleware.Middleware, error) {
 	cfg, err := decodeConfig(rawConfig)
 	if err != nil {
 		return nil, err
 	}
 
-	fctx := builtin.Context()
-	pricingPath := cfg.PricingPath
-	if pricingPath == "" {
-		pricingPath = defaultPricingFilename
+	if cfg.Pricing == nil {
+		if logger := builtin.Context().Logger; logger != nil {
+			logger.Warnf("cost_meter: no pricing table in middleware config; management predates config-delivered pricing — every request will record cost.skipped=unknown_model ($0)")
+		}
+		return newMiddleware(mustEmptyTable(), nil), nil
 	}
 
-	loader, err := pricing.NewLoader(fctx.DataDir, pricingPath, ID, nil)
+	defaults, err := pricing.NewTable(cfg.Pricing.Defaults)
 	if err != nil {
-		return nil, fmt.Errorf("init pricing loader: %w", err)
+		return nil, fmt.Errorf("cost_meter pricing defaults: %w", err)
 	}
-
-	cancel := startReloader(fctx.Context, loader)
-
-	return newMiddleware(loader, cancel), nil
+	perRecord, err := pricing.NewEntries(cfg.Pricing.Providers)
+	if err != nil {
+		return nil, fmt.Errorf("cost_meter per-provider pricing: %w", err)
+	}
+	return newMiddleware(defaults, perRecord), nil
 }
 
-// startReloader binds the loader's mtime-poll goroutine to a context
-// derived from the proxy-lifetime context and returns its cancel func so
-// the owning middleware can stop the goroutine on teardown. Returns nil
-// when there's nothing to watch (nil context or defaults-only loader), in
-// which case the middleware's Close is a no-op.
-func startReloader(ctx context.Context, loader *pricing.Loader) context.CancelFunc {
-	if ctx == nil || !loader.WatchesFile() {
-		return nil
+// mustEmptyTable returns a valid empty table. NewTable on a nil map cannot
+// fail; the panic guard documents that invariant.
+func mustEmptyTable() *pricing.Table {
+	t, err := pricing.NewTable(nil)
+	if err != nil {
+		panic(fmt.Sprintf("cost_meter: empty pricing table must build: %v", err))
 	}
-	cctx, cancel := context.WithCancel(ctx)
-	go loader.Reload(cctx)
-	return cancel
+	return t
 }
 
 // decodeConfig accepts empty, null, and {} configs, returning a

proxy/internal/middleware/builtin/cost_meter/middleware.go

@@ -1,7 +1,9 @@
 // Package cost_meter implements the SlotOnResponse middleware that
 // converts token-usage metadata emitted by llm_response_parser into a
-// per-request USD cost estimate. The middleware uses the shared pricing
-// loader so operator pricing overrides apply to the chain.
+// per-request USD cost estimate. Pricing arrives from management inside
+// the middleware config: a per-provider-record table (the operator's
+// stored prices, matched via llm.resolved_provider_id) consulted first,
+// then the surface-keyed defaults table.
 package cost_meter
 
 import (
@@ -17,7 +19,9 @@ import (
 const ID = "cost_meter"
 
 // Version is the implementation version emitted via the spec merge.
-const Version = "1.0.0"
+// 1.1.0: pricing is config-delivered (defaults + per-provider-record
+// entries) instead of proxy-embedded.
+const Version = "1.1.0"
 
 // Skip reasons emitted under KeyCostSkipped. The set is closed; the
 // dashboard surfaces these verbatim.
@@ -42,19 +46,21 @@ var metadataKeys = []string{
 }
 
 // Middleware computes a per-response cost estimate from the token
-// counts emitted upstream by llm_response_parser.
+// counts emitted upstream by llm_response_parser. Both tables are
+// immutable — a pricing change arrives as a mapping push that rebuilds
+// the chain with a fresh instance.
 type Middleware struct {
-	loader *pricing.Loader
-	// cancel stops this instance's pricing-reload goroutine. Non-nil only
-	// when the loader watches an override file; Close calls it so a chain
-	// rebuild doesn't leak a poll goroutine per retired instance.
-	cancel context.CancelFunc
+	// defaults is the surface-keyed table (llm.provider x llm.model).
+	defaults *pricing.Table
+	// perRecord is keyed by provider record id (llm.resolved_provider_id)
+	// then normalized model id; entries arrive fully materialized from
+	// management. Consulted before defaults. May be nil.
+	perRecord map[string]map[string]pricing.Entry
 }
 
-// newMiddleware constructs a Middleware bound to the given pricing loader.
-// cancel may be nil (defaults-only loader with no reloader to stop).
-func newMiddleware(loader *pricing.Loader, cancel context.CancelFunc) *Middleware {
-	return &Middleware{loader: loader, cancel: cancel}
+// newMiddleware constructs a Middleware over the given pricing tables.
+func newMiddleware(defaults *pricing.Table, perRecord map[string]map[string]pricing.Entry) *Middleware {
+	return &Middleware{defaults: defaults, perRecord: perRecord}
 }
 
 // ID returns the registry identifier.
@@ -79,16 +85,9 @@ func (m *Middleware) MetadataKeys() []string {
 // response.
 func (m *Middleware) MutationsSupported() bool { return false }
 
-// Close stops this instance's pricing-reload goroutine, if any. Called by
-// the chain when a rebuild retires the instance, so the mtime-poll loop
-// doesn't outlive the chain it belonged to. Safe to call on a nil receiver
-// and on an instance with no reloader.
-func (m *Middleware) Close() error {
-	if m != nil && m.cancel != nil {
-		m.cancel()
-	}
-	return nil
-}
+// Close releases resources owned by the middleware. Stateless — the
+// pricing tables are plain maps owned by this instance.
+func (m *Middleware) Close() error { return nil }
 
 // Invoke reads provider, model, and token metadata, looks up pricing,
 // and emits either KeyCostUSDTotal or KeyCostSkipped. The decision is
@@ -144,8 +143,7 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 		return out, nil
 	}
 
-	table := m.loader.Get()
-	costs, ok := table.Costs(provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
+	costs, ok := m.lookupCosts(in.Metadata, provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
 	if !ok {
 		out.Metadata = skip(skipUnknownModel)
 		return out, nil
@@ -164,6 +162,26 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 	return out, nil
 }
 
+// lookupCosts resolves the price for this request and computes the cost
+// split. Resolution order:
+//
+//  1. Per-provider-record entry: the operator's stored price for the
+//     provider route that served the request, keyed by the
+//     llm.resolved_provider_id metadata llm_router stamped on the allow
+//     path. Absent metadata (e.g. no router in the chain) skips this tier.
+//  2. Surface defaults: the catalog-derived table keyed by llm.provider.
+//
+// The surface always selects the cache formula — a per-record entry for an
+// Anthropic route still bills its cache buckets additively.
+func (m *Middleware) lookupCosts(md []middleware.KV, surface, model string, inTokens, outTokens, cachedTokens, cacheCreationTokens int64) (pricing.Costs, bool) {
+	if recordID := lookupKV(md, middleware.KeyLLMResolvedProviderID); recordID != "" {
+		if entry, ok := m.perRecord[recordID][model]; ok {
+			return pricing.EntryCosts(entry, surface, inTokens, outTokens, cachedTokens, cacheCreationTokens), true
+		}
+	}
+	return m.defaults.Costs(surface, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
+}
+
 // usd renders a cost as the fixed-precision string every cost.usd_* key
 // carries, so the per-bucket values and the aggregates round identically.
 //

proxy/internal/middleware/builtin/cost_meter/middleware_test.go

@@ -3,39 +3,50 @@ package cost_meter
 import (
 	"context"
 	"encoding/json"
-	"os"
-	"path/filepath"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 
+	"github.com/netbirdio/netbird/proxy/internal/llm/pricing"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
-	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 )
 
-const fixturePricing = `openai:
-  gpt-4o:
-    input_per_1k: 0.0025
-    output_per_1k: 0.01
-  gpt-4o-mini:
-    input_per_1k: 0.00015
-    output_per_1k: 0.0006
-anthropic:
-  claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-`
-
-// configureBuiltin points the package-level FactoryContext at a tmp
-// directory containing the test pricing fixture. Returns the path so
-// callers can override files later if needed.
-func configureBuiltin(t *testing.T) string {
+// fixtureConfig mirrors what management's buildCostMeterConfigJSON ships:
+// a surface-keyed defaults table. Rates match the retired YAML fixture so
+// every cost assertion below is byte-identical to the pre-feature values.
+func fixtureConfig(t *testing.T) []byte {
 	t.Helper()
-	dir := t.TempDir()
-	require.NoError(t, os.WriteFile(filepath.Join(dir, "pricing.yaml"), []byte(fixturePricing), 0o600), "write pricing fixture")
-	builtin.Configure(context.Background(), dir, nil, nil, nil)
-	return dir
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {
+				"gpt-4o":      {InputPer1K: 0.0025, OutputPer1K: 0.01},
+				"gpt-4o-mini": {InputPer1K: 0.00015, OutputPer1K: 0.0006},
+			},
+			"anthropic": {
+				"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015},
+			},
+		},
+	}})
+	require.NoError(t, err)
+	return raw
+}
+
+// fixtureConfigWithCache adds the cache-rate fields.
+func fixtureConfigWithCache(t *testing.T) []byte {
+	t.Helper()
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {
+				"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01, CachedInputPer1K: 0.00125},
+			},
+			"anthropic": {
+				"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015, CacheReadPer1K: 0.0003, CacheCreationPer1K: 0.00375},
+			},
+		},
+	}})
+	require.NoError(t, err)
+	return raw
 }
 
 func metaValue(t *testing.T, kvs []middleware.KV, key string) (string, bool) {
@@ -56,8 +67,7 @@ func buildMiddleware(t *testing.T, raw []byte) middleware.Middleware {
 }
 
 func TestMiddleware_StaticSurface(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	assert.Equal(t, ID, mw.ID(), "ID must match the registered constant")
 	assert.Equal(t, Version, mw.Version(), "Version must match the constant")
@@ -79,8 +89,10 @@ func TestMiddleware_StaticSurface(t *testing.T) {
 	assert.Equal(t, expected, keys, "metadata key allowlist must match the spec")
 }
 
+// TestFactory_AcceptsEmptyAndJSONConfig: empty/null/{} configs are what an
+// old management (pre config-delivered pricing) sends — they must build a
+// working (all-skip) instance, never fail the chain.
 func TestFactory_AcceptsEmptyAndJSONConfig(t *testing.T) {
-	configureBuiltin(t)
 	cases := [][]byte{nil, {}, []byte("null"), []byte("{}"), []byte("   ")}
 	for _, raw := range cases {
 		mw, err := Factory{}.New(raw)
@@ -90,15 +102,57 @@ func TestFactory_AcceptsEmptyAndJSONConfig(t *testing.T) {
 }
 
 func TestFactory_RejectsMalformedConfig(t *testing.T) {
-	configureBuiltin(t)
 	mw, err := Factory{}.New([]byte("{not json"))
 	require.Error(t, err, "malformed config must surface at construction")
 	assert.Nil(t, mw, "no instance is returned on error")
 }
 
-func TestFactory_DefaultPricingPathLoadsFixture(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+// TestFactory_RejectsInvalidRates: a non-finite or negative rate anywhere
+// in the table fails the chain build (defense-in-depth behind management's
+// API validation) rather than silently mispricing.
+func TestFactory_RejectsInvalidRates(t *testing.T) {
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {"gpt-4o": {InputPer1K: -0.0025, OutputPer1K: 0.01}},
+		},
+	}})
+	require.NoError(t, err)
+	mw, err := Factory{}.New(raw)
+	require.Error(t, err, "negative rate must fail the build")
+	assert.Nil(t, mw)
+
+	raw, err = json.Marshal(Config{Pricing: &PricingConfig{
+		Providers: map[string]map[string]pricing.EntryJSON{
+			"prov-1": {"m": {InputPer1K: 0.01, OutputPer1K: 0.01, CacheReadPer1K: -1}},
+		},
+	}})
+	require.NoError(t, err)
+	_, err = Factory{}.New(raw)
+	require.Error(t, err, "per-record tables validate too")
+}
+
+// TestFactory_NilPricingSkipsEverything is the version-skew contract: a
+// new proxy under an old management ({} config) must build, allow, and
+// skip with unknown_model — degraded but never broken.
+func TestFactory_NilPricingSkipsEverything(t *testing.T) {
+	mw := buildMiddleware(t, []byte("{}"))
+	out, err := mw.Invoke(context.Background(), &middleware.Input{
+		Metadata: []middleware.KV{
+			{Key: middleware.KeyLLMProvider, Value: "openai"},
+			{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+			{Key: middleware.KeyLLMInputTokens, Value: "1000"},
+			{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
+		},
+	})
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "cost_meter always allows")
+	value, ok := metaValue(t, out.Metadata, middleware.KeyCostSkipped)
+	require.True(t, ok, "no pricing table means every request skips")
+	assert.Equal(t, skipUnknownModel, value)
+}
+
+func TestFactory_ConfigDefaultsPriceRequests(t *testing.T) {
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -116,34 +170,78 @@ func TestFactory_DefaultPricingPathLoadsFixture(t *testing.T) {
 	assert.Equal(t, "0.000750000", value, "0.00015 + 0.0006 per 1k tokens, 9-decimal format")
 }
 
-func TestFactory_PricingPathOverride(t *testing.T) {
-	dir := t.TempDir()
-	require.NoError(t, os.WriteFile(filepath.Join(dir, "custom.yaml"), []byte(fixturePricing), 0o600), "write custom pricing")
-	builtin.Configure(context.Background(), dir, nil, nil, nil)
-
-	raw, err := json.Marshal(Config{PricingPath: "custom.yaml"})
+// TestInvoke_PerRecordEntryBeatsDefaults: when llm_router resolved a
+// provider record whose operator pinned a price for the model, that price
+// wins over the surface default.
+func TestInvoke_PerRecordEntryBeatsDefaults(t *testing.T) {
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01}},
+		},
+		Providers: map[string]map[string]pricing.EntryJSON{
+			"prov-azure": {"gpt-4o": {InputPer1K: 0.005, OutputPer1K: 0.02}},
+		},
+	}})
 	require.NoError(t, err)
-
 	mw := buildMiddleware(t, raw)
+
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMProvider, Value: "openai"},
 			{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
-			{Key: middleware.KeyLLMInputTokens, Value: "2000"},
+			{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-azure"},
+			{Key: middleware.KeyLLMInputTokens, Value: "1000"},
 			{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
 		},
 	})
 	require.NoError(t, err)
-
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
-	require.True(t, ok, "cost.usd_total must be emitted with custom pricing path")
-	assert.Equal(t, "0.015000000", value, "2*0.0025 + 1*0.01 = 0.015 with 9-decimal format")
+	require.True(t, ok)
+	assert.Equal(t, "0.025000000", value, "operator's per-record price (0.005+0.02) wins over the default (0.0025+0.01)")
 }
 
-func TestInvoke_ComputesCostForKnownModel(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+// TestInvoke_PerRecordMissFallsBackToDefaults: a resolved record with no
+// entry for this model (or no entries at all) falls through to the
+// surface defaults — gateway providers rely on exactly this.
+func TestInvoke_PerRecordMissFallsBackToDefaults(t *testing.T) {
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01}},
+		},
+		Providers: map[string]map[string]pricing.EntryJSON{
+			"prov-1": {"some-other-model": {InputPer1K: 1, OutputPer1K: 1}},
+		},
+	}})
+	require.NoError(t, err)
+	mw := buildMiddleware(t, raw)
 
+	for name, recordID := range map[string]string{
+		"record with other models": "prov-1",
+		"record with no entries":   "prov-gateway",
+	} {
+		t.Run(name, func(t *testing.T) {
+			out, err := mw.Invoke(context.Background(), &middleware.Input{
+				Metadata: []middleware.KV{
+					{Key: middleware.KeyLLMProvider, Value: "openai"},
+					{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+					{Key: middleware.KeyLLMResolvedProviderID, Value: recordID},
+					{Key: middleware.KeyLLMInputTokens, Value: "1000"},
+					{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
+				},
+			})
+			require.NoError(t, err)
+			value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
+			require.True(t, ok, "per-record miss must fall back to the surface default, not skip")
+			assert.Equal(t, "0.012500000", value, "default rates apply")
+		})
+	}
+}
+
+// TestInvoke_NoResolvedProviderIDUsesDefaults: metadata without a
+// resolved provider id (router denied, or a chain without llm_router)
+// prices from the defaults table directly.
+func TestInvoke_NoResolvedProviderIDUsesDefaults(t *testing.T) {
+	mw := buildMiddleware(t, fixtureConfig(t))
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMProvider, Value: "anthropic"},
@@ -153,17 +251,15 @@ func TestInvoke_ComputesCostForKnownModel(t *testing.T) {
 		},
 	})
 	require.NoError(t, err)
-
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
-	require.True(t, ok, "cost.usd_total must be emitted")
+	require.True(t, ok)
 	assert.Equal(t, "0.018000000", value, "0.003 + 0.015 = 0.018 with 9-decimal format")
 	_, skipped := metaValue(t, out.Metadata, middleware.KeyCostSkipped)
 	assert.False(t, skipped, "cost.skipped must not be set when cost is computed")
 }
 
 func TestInvoke_MissingProvider(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -179,8 +275,7 @@ func TestInvoke_MissingProvider(t *testing.T) {
 }
 
 func TestInvoke_MissingModel(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -196,8 +291,7 @@ func TestInvoke_MissingModel(t *testing.T) {
 }
 
 func TestInvoke_MissingTokens(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	cases := []struct {
 		name string
@@ -240,8 +334,7 @@ func TestInvoke_MissingTokens(t *testing.T) {
 }
 
 func TestInvoke_UnparseableTokens(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	cases := []struct {
 		name string
@@ -271,8 +364,7 @@ func TestInvoke_UnparseableTokens(t *testing.T) {
 }
 
 func TestInvoke_ZeroTokens(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -291,8 +383,7 @@ func TestInvoke_ZeroTokens(t *testing.T) {
 }
 
 func TestInvoke_UnknownModel(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -309,8 +400,7 @@ func TestInvoke_UnknownModel(t *testing.T) {
 }
 
 func TestInvoke_NilInput(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))
 
 	out, err := mw.Invoke(context.Background(), nil)
 	require.NoError(t, err)
@@ -319,36 +409,12 @@ func TestInvoke_NilInput(t *testing.T) {
 	assert.Empty(t, out.Metadata, "no metadata must be emitted on nil input")
 }
 
-const fixturePricingWithCache = `openai:
-  gpt-4o:
-    input_per_1k: 0.0025
-    output_per_1k: 0.01
-    cached_input_per_1k: 0.00125
-anthropic:
-  claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-`
-
-// configureBuiltinWithCacheRates points the package-level
-// FactoryContext at a tmp directory containing pricing entries that
-// include the cache rate fields.
-func configureBuiltinWithCacheRates(t *testing.T) {
-	t.Helper()
-	dir := t.TempDir()
-	require.NoError(t, os.WriteFile(filepath.Join(dir, "pricing.yaml"), []byte(fixturePricingWithCache), 0o600), "write cache-aware pricing fixture")
-	builtin.Configure(context.Background(), dir, nil, nil, nil)
-}
-
 // TestInvoke_OpenAICachedSubsetDiscount proves the OpenAI shape end
 // to end through the middleware: cached_input_tokens is treated as a
 // SUBSET of input_tokens and discounted at the configured rate, not
 // added on top.
 func TestInvoke_OpenAICachedSubsetDiscount(t *testing.T) {
-	configureBuiltinWithCacheRates(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfigWithCache(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -390,8 +456,7 @@ func TestInvoke_OpenAICachedSubsetDiscount(t *testing.T) {
 // shape: cache_read and cache_creation are additive to input_tokens
 // and each carries its own rate.
 func TestInvoke_AnthropicCacheBucketsAdditive(t *testing.T) {
-	configureBuiltinWithCacheRates(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfigWithCache(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -429,8 +494,37 @@ func TestInvoke_AnthropicCacheBucketsAdditive(t *testing.T) {
 		"output bucket bills 200 tokens at 0.015/1k")
 }
 
+// TestInvoke_PerRecordEntryUsesSurfaceFormula: a per-record entry for an
+// anthropic-surface request must bill its cache buckets additively — the
+// formula follows llm.provider, not which table the entry came from.
+func TestInvoke_PerRecordEntryUsesSurfaceFormula(t *testing.T) {
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Providers: map[string]map[string]pricing.EntryJSON{
+			"prov-ant": {"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015, CacheReadPer1K: 0.0003, CacheCreationPer1K: 0.00375}},
+		},
+	}})
+	require.NoError(t, err)
+	mw := buildMiddleware(t, raw)
+
+	out, err := mw.Invoke(context.Background(), &middleware.Input{
+		Metadata: []middleware.KV{
+			{Key: middleware.KeyLLMProvider, Value: "anthropic"},
+			{Key: middleware.KeyLLMModel, Value: "claude-sonnet-4-5"},
+			{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-ant"},
+			{Key: middleware.KeyLLMInputTokens, Value: "256"},
+			{Key: middleware.KeyLLMOutputTokens, Value: "200"},
+			{Key: middleware.KeyLLMCachedInputTokens, Value: "768"},
+			{Key: middleware.KeyLLMCacheCreationTokens, Value: "512"},
+		},
+	})
+	require.NoError(t, err)
+	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
+	require.True(t, ok)
+	assert.Equal(t, "0.005918400", value, "identical math to the defaults-table entry with the same rates")
+}
+
 // assertBucket asserts one per-bucket cost key carries the expected
-// 6-decimal value.
+// 9-decimal value.
 func assertBucket(t *testing.T, md []middleware.KV, key, want, msg string) {
 	t.Helper()
 	got, ok := metaValue(t, md, key)
@@ -439,13 +533,10 @@ func assertBucket(t *testing.T, md []middleware.KV, key, want, msg string) {
 }
 
 // TestInvoke_CachedTokensAbsentFallsBackToBaseFormula covers the
-// "operator hasn't opted in" path: with no cached metadata keys
-// emitted, the meter must produce exactly the same cost as before
-// the feature landed. Critical so operators with the new binary but
-// no YAML changes see no behavioural drift on OpenAI requests.
+// no-cache-metadata path: with no cached keys emitted, the meter must
+// produce exactly the input+output cost.
 func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
-	configureBuiltinWithCacheRates(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfigWithCache(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -460,7 +551,7 @@ func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok)
 	// 1000 input * 0.0025 + 500 output * 0.01 = 0.0025 + 0.005 = 0.0075
-	assert.Equal(t, "0.007500000", value, "no cached metadata = same cost as before the feature landed")
+	assert.Equal(t, "0.007500000", value, "no cached metadata = plain input+output cost")
 }
 
 // TestInvoke_UnparseableCachedTokensSkippedSilently proves the
@@ -469,8 +560,7 @@ func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
 // regular formula. Cache buckets are a refinement, never a reason to
 // abort cost computation.
 func TestInvoke_UnparseableCachedTokensSkippedSilently(t *testing.T) {
-	configureBuiltinWithCacheRates(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfigWithCache(t))
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -487,22 +577,10 @@ func TestInvoke_UnparseableCachedTokensSkippedSilently(t *testing.T) {
 	assert.Equal(t, "0.007500000", value, "same as the no-cached-metadata path")
 }
 
-// TestMiddleware_CloseCancelsReloader proves Close stops the per-instance
-// pricing-reload goroutine: a chain rebuild retires the old instance and
-// calls Close, which must invoke the cancel func startReloader handed it so
-// the mtime-poll loop doesn't outlive the chain.
-func TestMiddleware_CloseCancelsReloader(t *testing.T) {
-	ctx, cancel := context.WithCancel(context.Background())
-	m := newMiddleware(nil, cancel)
-
-	require.NoError(t, m.Close(), "Close must not error")
-	require.Error(t, ctx.Err(), "Close must cancel the reloader context so the poll goroutine exits")
-}
-
-// TestMiddleware_CloseNilSafe confirms Close is a no-op (no panic) for an
-// instance with no reloader and for a nil receiver.
+// TestMiddleware_CloseNilSafe confirms Close is a no-op (no panic) even
+// for a nil receiver.
 func TestMiddleware_CloseNilSafe(t *testing.T) {
-	require.NoError(t, newMiddleware(nil, nil).Close(), "no-reloader Close must be a no-op")
+	require.NoError(t, newMiddleware(nil, nil).Close(), "Close must be a no-op")
 	var m *Middleware
 	require.NoError(t, m.Close(), "nil-receiver Close must be safe")
 }

proxy/internal/middleware/builtin/llm_request_parser/bedrock_test.go

@@ -6,23 +6,6 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-func TestNormalizeBedrockModel(t *testing.T) {
-	cases := map[string]string{
-		"eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
-		"us.anthropic.claude-opus-4-8-20250101-v1:0":   "anthropic.claude-opus-4-8",
-		"apac.anthropic.claude-haiku-4-5-v1:0":         "anthropic.claude-haiku-4-5",
-		"anthropic.claude-sonnet-4-5-20250929-v1:0":    "anthropic.claude-sonnet-4-5",
-		"meta.llama3-3-70b-instruct-v1:0":              "meta.llama3-3-70b-instruct",
-		"amazon.nova-pro-v1:0":                         "amazon.nova-pro",
-		"amazon.nova-2-lite-v1:0":                      "amazon.nova-2-lite",
-		// Inference-profile ARN — model id lives in the last path segment.
-		"arn:aws:bedrock:eu-central-1:123456789012:inference-profile/eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
-	}
-	for in, want := range cases {
-		require.Equal(t, want, normalizeBedrockModel(in), "normalize %q", in)
-	}
-}
-
 func TestParseBedrockPath(t *testing.T) {
 	tests := []struct {
 		path   string

proxy/internal/middleware/builtin/llm_request_parser/middleware.go

@@ -8,7 +8,6 @@ package llm_request_parser
 import (
 	"context"
 	"net/url"
-	"regexp"
 	"strconv"
 	"strings"
 	"unicode/utf8"
@@ -253,9 +252,7 @@ func parseVertexPath(reqPath string) (vertexRequest, bool) {
 	if c := strings.LastIndex(rest, ":"); c >= 0 {
 		model, action = rest[:c], rest[c+1:]
 	}
-	if at := strings.Index(model, "@"); at >= 0 {
-		model = model[:at]
-	}
+	model = llm.NormalizeVertexModel(model)
 	if model == "" {
 		return vertexRequest{}, false
 	}
@@ -343,14 +340,6 @@ func trimBedrockNamespace(reqPath string) string {
 	return reqPath
 }
 
-// bedrockRegionPrefixes are the cross-region inference-profile prefixes that
-// front a Bedrock model id (e.g. "eu.anthropic.claude-...").
-var bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
-
-// bedrockVersionSuffix matches the trailing "-vN[:N]" or "-YYYYMMDD-vN[:N]"
-// version/throughput suffix of a Bedrock model id.
-var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
-
 // parseBedrockPath extracts the model and streaming/converse flags from an AWS
 // Bedrock runtime model endpoint:
 //
@@ -375,7 +364,7 @@ func parseBedrockPath(reqPath string) (bedrockRequest, bool) {
 	if decoded, err := url.PathUnescape(rawModel); err == nil {
 		rawModel = decoded
 	}
-	model := normalizeBedrockModel(rawModel)
+	model := llm.NormalizeBedrockModel(rawModel)
 	if model == "" {
 		return bedrockRequest{}, false
 	}
@@ -389,30 +378,6 @@ func parseBedrockPath(reqPath string) (bedrockRequest, bool) {
 	}
 }
 
-// normalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
-// prefix, and the version/throughput suffix from a Bedrock model id so it
-// matches the catalog/pricing key, e.g.
-// "eu.anthropic.claude-sonnet-4-5-20250929-v1:0" -> "anthropic.claude-sonnet-4-5"
-// and "arn:aws:bedrock:eu-central-1:123:inference-profile/eu.anthropic.claude-sonnet-4-5-20250929-v1:0"
-// -> "anthropic.claude-sonnet-4-5".
-func normalizeBedrockModel(modelID string) string {
-	m := modelID
-	// A full ARN (inference-profile / provisioned-throughput / foundation-model)
-	// carries the model id in its last path segment.
-	if strings.HasPrefix(m, "arn:") {
-		if i := strings.LastIndex(m, "/"); i >= 0 {
-			m = m[i+1:]
-		}
-	}
-	for _, p := range bedrockRegionPrefixes {
-		if strings.HasPrefix(m, p) {
-			m = m[len(p):]
-			break
-		}
-	}
-	return bedrockVersionSuffix.ReplaceAllString(m, "")
-}
-
 // invokeBedrock emits the model/provider/session/prompt for an AWS Bedrock
 // request. Bedrock is metered under the dedicated "bedrock" parser, which reads
 // both the InvokeModel and Converse response shapes.

proxy/internal/proxy/agent_network_chain_realstack_test.go

@@ -18,10 +18,10 @@ import (
 	"google.golang.org/grpc/credentials/insecure"
 	"google.golang.org/grpc/test/bufconn"
 
-	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
-	mgmtgrpc "github.com/netbirdio/netbird/management/internals/shared/grpc"
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
 	agentNetworkTypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
+	mgmtgrpc "github.com/netbirdio/netbird/management/internals/shared/grpc"
 	"github.com/netbirdio/netbird/management/server/store"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
 	"github.com/netbirdio/netbird/proxy/internal/middleware/bodytap"
@@ -134,14 +134,18 @@ func TestReverseProxy_AgentNetworkRequest_FullChain(t *testing.T) {
 		RedactPii:              true,
 	}))
 	require.NoError(t, st.SaveAgentNetworkProvider(ctx, &agentNetworkTypes.Provider{
-		ID:                providerID,
-		AccountID:         testAccountID,
-		ProviderID:        "openai_api",
-		Name:              "openai-fullchain-test",
-		UpstreamURL:       upstream.URL, // router rewrites to this
-		APIKey:            "sk-test",
-		Enabled:           true,
-		Models:            []agentNetworkTypes.ProviderModel{{ID: "gpt-5.4"}},
+		ID:          providerID,
+		AccountID:   testAccountID,
+		ProviderID:  "openai_api",
+		Name:        "openai-fullchain-test",
+		UpstreamURL: upstream.URL, // router rewrites to this
+		APIKey:      "sk-test",
+		Enabled:     true,
+		// Operator-pinned prices deliberately differ from the catalog's
+		// gpt-5.4 rates (0.0025/0.015) so the cost assertion below proves
+		// the per-provider-record price — not the default table — billed
+		// this request: stored price → synth → wire → cost_meter.
+		Models:            []agentNetworkTypes.ProviderModel{{ID: "gpt-5.4", InputPer1k: 0.004, OutputPer1k: 0.02}},
 		SessionPrivateKey: "priv",
 		SessionPublicKey:  "pub",
 	}))
@@ -176,7 +180,7 @@ func TestReverseProxy_AgentNetworkRequest_FullChain(t *testing.T) {
 
 	// ---- 5. Wire the middleware framework — same registry the proxy uses
 	// in production, configured with our bufconn-backed management client.
-	mwbuiltin.Configure(ctx, t.TempDir(), nil, testLogger, mgmtClient)
+	mwbuiltin.Configure(ctx, nil, testLogger, mgmtClient)
 	registry := mwbuiltin.DefaultRegistry()
 	mwMetrics, err := middleware.NewMetrics(nil)
 	require.NoError(t, err)
@@ -283,13 +287,23 @@ func TestReverseProxy_AgentNetworkRequest_FullChain(t *testing.T) {
 			if r.DimensionKind == agentNetworkTypes.DimensionGroup &&
 				r.DimensionID == adminGroupID &&
 				r.WindowSeconds == 60 &&
-				r.TokensInput+r.TokensOutput > 0 {
+				r.TokensInput+r.TokensOutput > 0 &&
+				r.CostUSD > 0 {
 				return true
 			}
 		}
 		return false
 	}, 5*time.Second, 50*time.Millisecond,
-		"Admins group consumption row must increment via the response leg — if this fails the proxy's respInput dropped UserGroups again or the parser/recorder wiring is broken")
+		"Admins group consumption row must increment via the response leg WITH a non-zero cost — a zero cost means the operator's stored price never reached cost_meter (synth → wire → per-record lookup broken)")
+
+	// 8a-cost. Exact cost from the OPERATOR's stored price, not the catalog
+	// default: 12 prompt tokens × 0.004/1k + 40 completion tokens × 0.02/1k
+	// = 0.000048 + 0.0008 = 0.000848. With catalog rates it would be 0.00063
+	// — this assertion distinguishes the two, closing the loop on the whole
+	// dynamic-pricing feature (dashboard save → synth → gRPC wire →
+	// llm.resolved_provider_id lookup → billing).
+	assert.Equal(t, "0.000848000", cd.GetMetadata()["cost.usd_total"],
+		"cost must be computed from the provider record's operator-pinned price")
 
 	// 8b. Both the captured prompt and the captured completion are
 	// redacted — proves the synth threads redact_pii=true into BOTH parser

proxy/server.go

@@ -246,10 +246,6 @@ type Server struct {
 	// in processMappings before the receive loop reconnects to resync.
 	// Zero uses defaultMappingBatchWatchdog.
 	MappingBatchWatchdog time.Duration
-	// MiddlewareDataDir is the base directory the middleware system uses to
-	// resolve file-backed configuration (e.g. the cost_meter pricing table).
-	// Empty means any middleware that requires a file fails at configure time.
-	MiddlewareDataDir string
 	// MiddlewareCaptureBudgetBytes overrides the proxy-wide in-flight capture
 	// budget passed to middleware.NewManager. Zero or negative values fall
 	// back to defaultMiddlewareCaptureBudgetBytes (256 MiB).
@@ -2093,7 +2089,7 @@ func (s *Server) initMiddlewareManager(ctx context.Context) error {
 		return fmt.Errorf("middleware manager requires metrics bundle")
 	}
 	otelMeter := s.meter.Meter()
-	mwbuiltin.Configure(ctx, s.MiddlewareDataDir, otelMeter, s.Logger, s.mgmtClient)
+	mwbuiltin.Configure(ctx, otelMeter, s.Logger, s.mgmtClient)
 
 	mwMetrics, err := middleware.NewMetrics(otelMeter)
 	if err != nil {

shared/llm/model.go

@@ -0,0 +1,58 @@
+// Package llm holds LLM model-identifier helpers shared by the proxy and
+// the management server. The proxy normalizes model ids parsed off inbound
+// requests; management normalizes the operator's registered model ids at
+// synthesis time so both sides of the pricing / routing contract compare
+// equal.
+package llm
+
+import (
+	"regexp"
+	"strings"
+)
+
+// bedrockRegionPrefixes are the cross-region inference-profile prefixes that
+// front a Bedrock model id (e.g. "eu.anthropic.claude-...").
+var bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
+
+// bedrockVersionSuffix matches the trailing "-vN[:N]" or "-YYYYMMDD-vN[:N]"
+// version/throughput suffix of a Bedrock model id.
+var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
+
+// NormalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
+// prefix, and the version/throughput suffix from a Bedrock model id so it
+// matches the catalog/pricing key, e.g.
+// "eu.anthropic.claude-sonnet-4-5-20250929-v1:0" -> "anthropic.claude-sonnet-4-5"
+// and the inference-profile ARN's last segment likewise. It is the single
+// source of truth shared by the proxy's request parser (which normalizes the
+// request model from the URL path), the proxy's router (which normalizes the
+// operator's registered Bedrock model ids so both sides compare equal), and
+// the management synthesizer (which keys per-provider pricing entries by the
+// normalized id the parser will emit at billing time).
+func NormalizeBedrockModel(modelID string) string {
+	m := modelID
+	// A full ARN (inference-profile / provisioned-throughput / foundation-model)
+	// carries the model id in its last path segment.
+	if strings.HasPrefix(m, "arn:") {
+		if i := strings.LastIndex(m, "/"); i >= 0 {
+			m = m[i+1:]
+		}
+	}
+	for _, p := range bedrockRegionPrefixes {
+		if strings.HasPrefix(m, p) {
+			m = m[len(p):]
+			break
+		}
+	}
+	return bedrockVersionSuffix.ReplaceAllString(m, "")
+}
+
+// NormalizeVertexModel strips the "@version" suffix from a Vertex AI model id
+// (e.g. "claude-sonnet-4-5@20250929" -> "claude-sonnet-4-5") so it matches
+// the catalog/pricing key. Vertex publisher models are priced under their
+// vendor surface with the bare, unversioned id.
+func NormalizeVertexModel(modelID string) string {
+	if at := strings.Index(modelID, "@"); at >= 0 {
+		return modelID[:at]
+	}
+	return modelID
+}

shared/llm/model_test.go

@@ -11,6 +11,8 @@ func TestNormalizeBedrockModel(t *testing.T) {
 		"eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
 		"us.anthropic.claude-haiku-4-5":                "anthropic.claude-haiku-4-5",
 		"us.anthropic.claude-opus-4-8-20250101-v1:0":   "anthropic.claude-opus-4-8",
+		"apac.anthropic.claude-haiku-4-5-v1:0":         "anthropic.claude-haiku-4-5",
+		"amazon.nova-2-lite-v1:0":                      "amazon.nova-2-lite",
 		"anthropic.claude-sonnet-4-5-20250929-v1:0":    "anthropic.claude-sonnet-4-5",
 		"meta.llama3-3-70b-instruct-v1:0":              "meta.llama3-3-70b-instruct",
 		"amazon.nova-pro-v1:0":                         "amazon.nova-pro",
@@ -21,3 +23,14 @@ func TestNormalizeBedrockModel(t *testing.T) {
 		require.Equal(t, want, NormalizeBedrockModel(in), "normalize %q", in)
 	}
 }
+
+func TestNormalizeVertexModel(t *testing.T) {
+	cases := map[string]string{
+		"claude-sonnet-4-5@20250929": "claude-sonnet-4-5",
+		"claude-haiku-4-5":           "claude-haiku-4-5",
+		"gpt-4o@2024-08-06":          "gpt-4o",
+	}
+	for in, want := range cases {
+		require.Equal(t, want, NormalizeVertexModel(in), "normalize %q", in)
+	}
+}

shared/management/http/api/openapi.yml

@@ -5271,6 +5271,21 @@ components:
           format: double
           description: Cost per 1k output tokens, in USD.
           example: 0.0006
+        cached_input_per_1k:
+          type: number
+          format: double
+          description: OpenAI-shape cache rate — cost per 1k cached prompt tokens (a subset of input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means no discount (cached tokens bill at input_per_1k).
+          example: 0.000075
+        cache_read_per_1k:
+          type: number
+          format: double
+          description: Anthropic-shape cache rate — cost per 1k cache-read tokens (additive to input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means cache reads bill at input_per_1k.
+          example: 0.0003
+        cache_creation_per_1k:
+          type: number
+          format: double
+          description: Anthropic-shape cache rate — cost per 1k cache-creation tokens (additive to input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means cache writes bill at input_per_1k.
+          example: 0.00375
       required:
         - id
         - input_per_1k
@@ -5296,6 +5311,21 @@ components:
           format: double
           description: Output token price per 1k tokens, in USD.
           example: 0.015
+        cached_input_per_1k:
+          type: number
+          format: double
+          description: OpenAI-shape cache rate — default cost per 1k cached prompt tokens (a subset of input tokens), in USD. Absent when the model has no cached-input discount.
+          example: 0.000075
+        cache_read_per_1k:
+          type: number
+          format: double
+          description: Anthropic-shape cache rate — default cost per 1k cache-read tokens (additive to input tokens), in USD. Absent when the model has no cache-read rate.
+          example: 0.0003
+        cache_creation_per_1k:
+          type: number
+          format: double
+          description: Anthropic-shape cache rate — default cost per 1k cache-creation tokens (additive to input tokens), in USD. Absent when the model has no cache-creation rate.
+          example: 0.00375
         context_window:
           type: integer
           description: Maximum context window in tokens.
@@ -5354,6 +5384,13 @@ components:
             $ref: '#/components/schemas/AgentNetworkCatalogExtraHeader'
         identity_injection:
           $ref: '#/components/schemas/AgentNetworkCatalogIdentityInjection'
+        pricing_surfaces:
+          type: array
+          description: |
+            Cost-meter pricing surfaces this provider's traffic is metered under ("openai", "anthropic", "bedrock"). Tells the dashboard which cache-rate fields apply to this provider's models: "openai" → cached_input_per_1k (cached prompt tokens are a subset of input); "anthropic"/"bedrock" → cache_read_per_1k + cache_creation_per_1k (additive buckets). Absent/empty for gateway and custom entries, whose upstream shape NetBird cannot know ahead of time — surface all cache fields for those.
+          items:
+            type: string
+          example: ["openai"]
         models:
           type: array
           description: Catalog models available for this provider.

shared/management/http/api/types.gen.go

@@ -2017,6 +2017,15 @@ type AgentNetworkCatalogJSONMetadataInjection struct {
 
 // AgentNetworkCatalogModel defines model for AgentNetworkCatalogModel.
 type AgentNetworkCatalogModel struct {
+	// CacheCreationPer1k Anthropic-shape cache rate — default cost per 1k cache-creation tokens (additive to input tokens), in USD. Absent when the model has no cache-creation rate.
+	CacheCreationPer1k *float64 `json:"cache_creation_per_1k,omitempty"`
+
+	// CacheReadPer1k Anthropic-shape cache rate — default cost per 1k cache-read tokens (additive to input tokens), in USD. Absent when the model has no cache-read rate.
+	CacheReadPer1k *float64 `json:"cache_read_per_1k,omitempty"`
+
+	// CachedInputPer1k OpenAI-shape cache rate — default cost per 1k cached prompt tokens (a subset of input tokens), in USD. Absent when the model has no cached-input discount.
+	CachedInputPer1k *float64 `json:"cached_input_per_1k,omitempty"`
+
 	// ContextWindow Maximum context window in tokens.
 	ContextWindow int `json:"context_window"`
 
@@ -2070,6 +2079,9 @@ type AgentNetworkCatalogProvider struct {
 
 	// Name Display name for the provider.
 	Name string `json:"name"`
+
+	// PricingSurfaces Cost-meter pricing surfaces this provider's traffic is metered under ("openai", "anthropic", "bedrock"). Tells the dashboard which cache-rate fields apply to this provider's models: "openai" → cached_input_per_1k (cached prompt tokens are a subset of input); "anthropic"/"bedrock" → cache_read_per_1k + cache_creation_per_1k (additive buckets). Absent/empty for gateway and custom entries, whose upstream shape NetBird cannot know ahead of time — surface all cache fields for those.
+	PricingSurfaces *[]string `json:"pricing_surfaces,omitempty"`
 }
 
 // AgentNetworkCatalogProviderKind Presentation grouping for the provider Select on the dashboard.
@@ -2293,6 +2305,15 @@ type AgentNetworkProvider struct {
 
 // AgentNetworkProviderModel A model exposed by the provider, with the operator's per-1k input/output prices in USD.
 type AgentNetworkProviderModel struct {
+	// CacheCreationPer1k Anthropic-shape cache rate — cost per 1k cache-creation tokens (additive to input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means cache writes bill at input_per_1k.
+	CacheCreationPer1k *float64 `json:"cache_creation_per_1k,omitempty"`
+
+	// CacheReadPer1k Anthropic-shape cache rate — cost per 1k cache-read tokens (additive to input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means cache reads bill at input_per_1k.
+	CacheReadPer1k *float64 `json:"cache_read_per_1k,omitempty"`
+
+	// CachedInputPer1k OpenAI-shape cache rate — cost per 1k cached prompt tokens (a subset of input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means no discount (cached tokens bill at input_per_1k).
+	CachedInputPer1k *float64 `json:"cached_input_per_1k,omitempty"`
+
 	// Id Model identifier (e.g. "gpt-4o-mini").
 	Id string `json:"id"`
 

shared/management/types/network.go

@@ -47,6 +47,10 @@ type NetworkMap struct {
 	ForwardingRules     []*ForwardingRule
 	AuthorizedUsers     map[string]map[string]struct{}
 	EnableSSH           bool
+	// ForceRoutingPeerDNSResolution forces the peer to run/use routing-peer DNS
+	// resolution regardless of the account-global setting, for reverse-proxy
+	// domain targets.
+	ForceRoutingPeerDNSResolution bool
 }
 
 func (nm *NetworkMap) Merge(other *NetworkMap) {
@@ -56,6 +60,7 @@ func (nm *NetworkMap) Merge(other *NetworkMap) {
 	nm.FirewallRules = mergeUnique(nm.FirewallRules, other.FirewallRules)
 	nm.RoutesFirewallRules = mergeUnique(nm.RoutesFirewallRules, other.RoutesFirewallRules)
 	nm.ForwardingRules = mergeUnique(nm.ForwardingRules, other.ForwardingRules)
+	nm.ForceRoutingPeerDNSResolution = nm.ForceRoutingPeerDNSResolution || other.ForceRoutingPeerDNSResolution
 }
 
 type comparableObject[T any] interface {

shared/management/types/networkmap_components.go

@@ -56,6 +56,11 @@ type NetworkMapComponents struct {
 
 	// true when returning an empty-like map (returned instead of nil)
 	empty bool
+
+	// ForceRoutingPeerDNSResolution forces the peer to run/use routing-peer DNS
+	// resolution regardless of the account-global setting, for reverse-proxy
+	// domain targets.
+	ForceRoutingPeerDNSResolution bool
 }
 
 type routeIndexEntry struct {
@@ -190,6 +195,8 @@ func (c *NetworkMapComponents) Calculate(ctx context.Context) *NetworkMap {
 		RoutesFirewallRules: append(networkResourcesFirewallRules, routesFirewallRules...),
 		AuthorizedUsers:     authorizedUsers,
 		EnableSSH:           sshEnabled,
+
+		ForceRoutingPeerDNSResolution: c.ForceRoutingPeerDNSResolution,
 	}
 }