Merge branch 'main' into reverse-proxy-allow-match-or

# Conflicts: # management/server/store/sql_store.go # management/server/store/sql_store_service_test.go
2026-10-13 00:49:07 +02:00 · 2026-07-31 21:57:57 +02:00
parent 2beb90aeeb 0780a806f2
commit f05f3fc005
216 changed files with 13000 additions and 4032 deletions
@@ -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 {
@@ -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
 }
@@ -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)
+	}
+}
@@ -86,12 +86,17 @@ type Manager interface {

 // PolicySelectionInput is the per-request selection envelope. The
 // proxy populates it from CapturedData (account, user, groups) plus
-// the provider llm_router resolved.
+// the provider llm_router resolved and the model it extracted.
 type PolicySelectionInput struct {
 	AccountID  string
 	UserID     string
 	GroupIDs   []string
 	ProviderID string
+	// Model is the already-normalised upstream model id the proxy extracted
+	// (parser strips Bedrock region/version, Vertex @version), so a
+	// case-insensitive compare suffices. Empty = undetermined → not permitted
+	// (fail closed).
+	Model string
 }

 // PolicySelectionResult names the policy that "pays" for this request
@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"math"
 	"sort"
+	"strings"
 	"time"

 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
@@ -35,6 +36,10 @@ const (
 	denyCodeAccountTokenCapExceeded = "llm_account.token_cap_exceeded"
 	//nolint:gosec // account deny code label, not a credential
 	denyCodeAccountBudgetCapExceeded = "llm_account.budget_cap_exceeded"
+	// denyCodeModelBlocked is returned when policies govern the request's
+	// (provider, caller-groups) but none permits the model. Matches the proxy
+	// guardrail's code so both layers surface the same label.
+	denyCodeModelBlocked = "llm_policy.model_blocked"
 )

 // consumptionCache holds the consumption counters prefetched for one
@@ -159,6 +164,25 @@ func (m *managerImpl) SelectPolicyForRequest(ctx context.Context, in PolicySelec
 	}
 	candidates := filterApplicablePolicies(policies, in)

+	// Model-allowlist gate scoped to the matched policies: keep candidates whose
+	// guardrails permit the model (none enabled = unrestricted), deny when
+	// policies apply but none permits it. Skip the load when none has a guardrail.
+	if len(candidates) > 0 && anyPolicyHasGuardrails(candidates) {
+		guardrailsByID, gErr := m.loadGuardrailsByID(ctx, in.AccountID)
+		if gErr != nil {
+			return nil, gErr
+		}
+		permitted := filterModelPermittedPolicies(candidates, guardrailsByID, in.Model)
+		if len(permitted) == 0 {
+			return &PolicySelectionResult{
+				Allow:      false,
+				DenyCode:   denyCodeModelBlocked,
+				DenyReason: modelBlockedReason(in.Model),
+			}, nil
+		}
+		candidates = permitted
+	}
+
 	// Prefetch every consumption counter the ceiling + candidate policies will
 	// read, in a single store round-trip, then score against the cache.
 	cache, err := m.prefetchConsumption(ctx, in, rules, candidates, now)
@@ -250,6 +274,90 @@ func filterApplicablePolicies(policies []*types.Policy, in PolicySelectionInput)
 	return out
 }

+// anyPolicyHasGuardrails reports whether any policy references at least one
+// guardrail, so the selector can skip loading guardrails when none do.
+func anyPolicyHasGuardrails(policies []*types.Policy) bool {
+	for _, p := range policies {
+		if p != nil && len(p.GuardrailIDs) > 0 {
+			return true
+		}
+	}
+	return false
+}
+
+// loadGuardrailsByID loads the account's guardrails indexed by ID. Used by the
+// model-allowlist gate to resolve each candidate policy's attached guardrails.
+func (m *managerImpl) loadGuardrailsByID(ctx context.Context, accountID string) (map[string]*types.Guardrail, error) {
+	guardrails, err := m.store.GetAccountAgentNetworkGuardrails(ctx, store.LockingStrengthNone, accountID)
+	if err != nil {
+		return nil, fmt.Errorf("list account guardrails: %w", err)
+	}
+	byID := make(map[string]*types.Guardrail, len(guardrails))
+	for _, g := range guardrails {
+		if g != nil {
+			byID[g.ID] = g
+		}
+	}
+	return byID, nil
+}
+
+// filterModelPermittedPolicies returns the subset of policies whose guardrails
+// permit the model. Order is preserved so downstream scoring is unaffected.
+func filterModelPermittedPolicies(policies []*types.Policy, byID map[string]*types.Guardrail, model string) []*types.Policy {
+	out := make([]*types.Policy, 0, len(policies))
+	for _, p := range policies {
+		if policyPermitsModel(p, byID, model) {
+			out = append(out, p)
+		}
+	}
+	return out
+}
+
+// policyPermitsModel reports whether a policy permits the model. No
+// allowlist-enabled guardrail = unrestricted (permits any, incl. empty);
+// otherwise the model must be in the union of its allowlists, so an
+// empty/undetermined model fails closed.
+func policyPermitsModel(p *types.Policy, byID map[string]*types.Guardrail, model string) bool {
+	if p == nil {
+		return false
+	}
+	wanted := normaliseModelID(model)
+	restricted := false
+	for _, gID := range p.GuardrailIDs {
+		g, ok := byID[gID]
+		if !ok || g == nil || !g.Checks.ModelAllowlist.Enabled {
+			continue
+		}
+		restricted = true
+		if wanted == "" {
+			continue
+		}
+		for _, allowed := range g.Checks.ModelAllowlist.Models {
+			if normaliseModelID(allowed) == wanted {
+				return true
+			}
+		}
+	}
+	return !restricted
+}
+
+// normaliseModelID lowercases and trims a model identifier so the allowlist
+// compare is case-insensitive and trim-tolerant. Mirrors the proxy guardrail's
+// normaliseModel so both layers agree on what "same model" means.
+func normaliseModelID(model string) string {
+	return strings.ToLower(strings.TrimSpace(model))
+}
+
+// modelBlockedReason builds the human-readable deny reason for a model-allowlist
+// rejection. The model is quoted when known; an undetermined model is reported
+// as such so the access log distinguishes "wrong model" from "no model".
+func modelBlockedReason(model string) string {
+	if normaliseModelID(model) == "" {
+		return "request model could not be determined for the policy allowlist"
+	}
+	return fmt.Sprintf("model %q is not permitted by any applicable policy allowlist", model)
+}
+
 // candidate is the per-policy intermediate the selector ranks. A
 // policy that's been exhausted on any enabled cap never makes it
 // into this slice; the selector's deny envelope carries the latest
@@ -0,0 +1,329 @@
+package agentnetwork
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"github.com/golang/mock/gomock"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+	"github.com/netbirdio/netbird/management/server/store"
+)
+
+// guardedPolicy builds an enabled, uncapped policy that authorises sourceGroups
+// to reach providerID under the given guardrails. Uncapped keeps the selector's
+// headroom scoring trivial so these tests isolate the model-allowlist gate.
+func guardedPolicy(id, account string, sourceGroups []string, providerID string, guardrailIDs ...string) *types.Policy {
+	return &types.Policy{
+		ID:                     id,
+		AccountID:              account,
+		Enabled:                true,
+		SourceGroups:           sourceGroups,
+		DestinationProviderIDs: []string{providerID},
+		GuardrailIDs:           guardrailIDs,
+		CreatedAt:              time.Now().UTC(),
+	}
+}
+
+// allowlistGuardrail builds a guardrail whose model allowlist is enabled and
+// carries the given models.
+func allowlistGuardrail(id, account string, models ...string) *types.Guardrail {
+	return &types.Guardrail{
+		ID:        id,
+		AccountID: account,
+		Checks: types.GuardrailChecks{
+			ModelAllowlist: types.GuardrailModelAllowlist{Enabled: true, Models: models},
+		},
+	}
+}
+
+func expectPolicies(mockStore *store.MockStore, account string, policies ...*types.Policy) {
+	mockStore.EXPECT().
+		GetAccountAgentNetworkPolicies(gomock.Any(), gomock.Any(), account).
+		Return(policies, nil)
+}
+
+func expectGuardrails(mockStore *store.MockStore, account string, guardrails ...*types.Guardrail) {
+	mockStore.EXPECT().
+		GetAccountAgentNetworkGuardrails(gomock.Any(), gomock.Any(), account).
+		Return(guardrails, nil)
+}
+
+// TestSelectPolicy_ModelBlockedByAllowlist proves the authoritative allowlist
+// decision: a policy authorises the (provider, group) but restricts the model,
+// and the requested model isn't on the list, so the request is denied.
+func TestSelectPolicy_ModelBlockedByAllowlist(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		UserID:     "user-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "a model outside the only applicable policy's allowlist must be denied")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode, "deny code must be model_blocked")
+	assert.NotEmpty(t, res.DenyReason, "deny reason must be populated")
+}
+
+// TestSelectPolicy_ModelAllowedByAllowlist is the allow counterpart: the model
+// is on the applicable policy's allowlist, so selection proceeds normally.
+func TestSelectPolicy_ModelAllowedByAllowlist(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o", "claude-opus-4"))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		UserID:     "user-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "a model on the applicable policy's allowlist must be allowed")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_CaseInsensitiveModelMatch proves the compare tolerates case
+// and surrounding whitespace, matching the proxy guardrail's normalisation.
+func TestSelectPolicy_CaseInsensitiveModelMatch(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "  GPT-4o  "))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "gpt-4o",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "case/whitespace variants must match the allowlist entry")
+}
+
+// TestSelectPolicy_UnguardedPolicyIsUnrestricted is the false-deny fix: when two
+// policies authorise the same (provider, group) and one has no guardrail, that
+// policy makes the request unrestricted — not caught by the other's allowlist.
+func TestSelectPolicy_UnguardedPolicyIsUnrestricted(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	restricted := guardedPolicy("pol-restricted", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	open := guardedPolicy("pol-open", "acc-1", []string{"grp-eng"}, "prov-1") // no guardrail
+	expectPolicies(mockStore, "acc-1", restricted, open)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "an un-guardrailed policy for the same (provider, group) must leave the request unrestricted")
+	assert.Equal(t, "pol-open", res.SelectedPolicyID, "the unrestricted policy must be the one that pays")
+}
+
+// TestSelectPolicy_AllowlistDoesNotLeakAcrossGroups is the false-allow fix: a
+// model allowlisted only for grp-b must not be usable by a grp-a caller. The
+// selector considers only policies applicable to the caller's groups.
+func TestSelectPolicy_AllowlistDoesNotLeakAcrossGroups(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	polA := guardedPolicy("pol-a", "acc-1", []string{"grp-a"}, "prov-1", "g-a")
+	polB := guardedPolicy("pol-b", "acc-1", []string{"grp-b"}, "prov-1", "g-b")
+	expectPolicies(mockStore, "acc-1", polA, polB)
+	expectGuardrails(mockStore, "acc-1",
+		allowlistGuardrail("g-a", "acc-1", "gpt-4o"),
+		allowlistGuardrail("g-b", "acc-1", "claude-opus-4"),
+	)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-a"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4", // only allowed for grp-b
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "grp-b's allowlisted model must not leak to a grp-a caller")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode)
+}
+
+// TestSelectPolicy_UndeterminedModelFailsClosed proves the fail-closed contract
+// mirrors the proxy: with a restricted applicable policy and an empty model
+// (e.g. a path-routed shape the parser couldn't map), the request is denied.
+func TestSelectPolicy_UndeterminedModelFailsClosed(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "", // undetermined
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "an undetermined model must fail closed against a restricted policy")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode)
+}
+
+// TestSelectPolicy_DisabledAllowlistDoesNotRestrict proves a guardrail whose
+// model allowlist is disabled imposes no model restriction, even though the
+// policy references it.
+func TestSelectPolicy_DisabledAllowlistDoesNotRestrict(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	disabled := &types.Guardrail{
+		ID:        "g-1",
+		AccountID: "acc-1",
+		Checks: types.GuardrailChecks{
+			ModelAllowlist: types.GuardrailModelAllowlist{Enabled: false, Models: []string{"gpt-4o"}},
+		},
+	}
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", disabled)
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "anything-goes",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "a disabled allowlist must not restrict the model")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_UnionAcrossPolicyGuardrails proves a policy with multiple
+// allowlist guardrails permits the union of their models (not just the first).
+func TestSelectPolicy_UnionAcrossPolicyGuardrails(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1", "g-2")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1",
+		allowlistGuardrail("g-1", "acc-1", "gpt-4o"),
+		allowlistGuardrail("g-2", "acc-1", "claude-opus-4"),
+	)
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4", // only in the second guardrail's list
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "a model in any of the policy's allowlist guardrails must be permitted")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_GuardrailLookupErrorPropagates proves a store failure while
+// resolving the candidate policies' guardrails surfaces as an error, not a
+// silent allow/deny.
+func TestSelectPolicy_GuardrailLookupErrorPropagates(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	mockStore.EXPECT().
+		GetAccountAgentNetworkGuardrails(gomock.Any(), gomock.Any(), "acc-1").
+		Return(nil, errors.New("store unavailable"))
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "gpt-4o",
+	})
+	require.Error(t, err, "a guardrail-lookup failure must surface as an error")
+	assert.Nil(t, res)
+}
+
+// TestSelectPolicy_MissingGuardrailReferenceTreatedAsUnrestricted proves a
+// policy referencing a guardrail ID absent from the account's set (a stale
+// reference) imposes no model restriction — same as no guardrail.
+func TestSelectPolicy_MissingGuardrailReferenceTreatedAsUnrestricted(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-missing")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1")
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "anything-goes",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "an orphaned guardrail reference must not restrict the model")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_PartialCandidatesPermittedAfterModelFilter proves the model
+// gate narrows candidates before cap scoring: the permitting policy is selected
+// even though the blocked one has a larger, more attractive cap.
+func TestSelectPolicy_PartialCandidatesPermittedAfterModelFilter(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	polBig := guardedPolicy("pol-big", "acc-1", []string{"grp-eng"}, "prov-1", "g-restrict")
+	polBig.Limits = types.PolicyLimits{
+		TokenLimit: types.PolicyTokenLimit{Enabled: true, GroupCap: 1_000_000, WindowSeconds: 3600},
+	}
+	polSmall := guardedPolicy("pol-small", "acc-1", []string{"grp-eng"}, "prov-1", "g-permit")
+	polSmall.Limits = types.PolicyLimits{
+		TokenLimit: types.PolicyTokenLimit{Enabled: true, GroupCap: 100, WindowSeconds: 3600},
+	}
+	expectPolicies(mockStore, "acc-1", polBig, polSmall)
+	expectGuardrails(mockStore, "acc-1",
+		allowlistGuardrail("g-restrict", "acc-1", "gpt-4o"),
+		allowlistGuardrail("g-permit", "acc-1", "claude-opus-4"),
+	)
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4", // only pol-small's guardrail permits this
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow)
+	assert.Equal(t, "pol-small", res.SelectedPolicyID,
+		"the model filter must exclude pol-big before cap scoring")
+}
@@ -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
+}
@@ -0,0 +1,282 @@
+# 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.
+
+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:
+  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
+  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
+    cached_input_per_1k: 0.0005
+  gpt-4.1-mini:
+    input_per_1k: 0.0004
+    output_per_1k: 0.0016
+    cached_input_per_1k: 0.0001
+  gpt-4.1-nano:
+    input_per_1k: 0.0001
+    output_per_1k: 0.0004
+    cached_input_per_1k: 0.000025
+  gpt-4o:
+    input_per_1k: 0.0025
+    output_per_1k: 0.01
+    cached_input_per_1k: 0.00125
+  gpt-4o-mini:
+    input_per_1k: 0.00015
+    output_per_1k: 0.0006
+    cached_input_per_1k: 0.000075
+  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
+    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
+  ministral-3-14b-2512:
+    input_per_1k: 0.0002
+    output_per_1k: 0.0002
+  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
+  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: 0.00018
+  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
@@ -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")
+}
@@ -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.
+`
@@ -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)
+	}
+}
@@ -0,0 +1,249 @@
+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 {
+			log.Infof("agent-network pricing defaults file %s not present; serving built-in defaults", path)
+			return nil
+		}
+		return err
+	}
+	storeFileTable(table, mtime)
+	log.Infof("agent-network pricing defaults loaded from %s", path)
+	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()
+
+	log.Debugf("agent-network pricing defaults reload: checking %s for changes", path)
+
+	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 {
+		log.Debugf("agent-network pricing defaults %s unchanged since last check", path)
+		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
+}
@@ -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")
+}
@@ -233,9 +233,19 @@ 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)
-	guardrailJSON, err := marshalGuardrailConfig(mergedGuardrails)
+	// The proxy guardrail is a per-provider fail-closed backstop; the
+	// authoritative per-policy/group decision is management's
+	// SelectPolicyForRequest. A provider lands in this map only when every
+	// authorising policy restricts models.
+	providerAllowlists := buildProviderAllowlists(enabledPolicies, guardrailsByID)
+	guardrailJSON, err := marshalGuardrailConfig(providerAllowlists, mergedGuardrails.PromptCapture)
 	if err != nil {
 		return nil, err
 	}
@@ -243,7 +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 {
@@ -695,7 +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
@@ -764,10 +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,
@@ -780,10 +793,12 @@ func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON []byt

 // guardrailConfig is the JSON shape the proxy-side llm_guardrail
 // middleware expects. Mirrors the proxy registration documented in
-// the management→proxy contract.
+// the management→proxy contract. provider_allowlists is keyed by the
+// resolved provider id llm_router stamps; a provider absent from the map is
+// unrestricted at the proxy layer.
 type guardrailConfig struct {
-	ModelAllowlist []string               `json:"model_allowlist,omitempty"`
-	PromptCapture  guardrailPromptCapture `json:"prompt_capture"`
+	ProviderAllowlists map[string][]string    `json:"provider_allowlists,omitempty"`
+	PromptCapture      guardrailPromptCapture `json:"prompt_capture"`
 }

 type guardrailPromptCapture struct {
@@ -828,13 +843,10 @@ func applyAccountCollectionControls(merged *MergedGuardrails, settings *types.Se
 	merged.PromptCapture.RedactPii = settings.RedactPii || merged.PromptCapture.RedactPii
 }

-func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
+func marshalGuardrailConfig(providerAllowlists map[string][]string, capture MergedPromptCapture) ([]byte, error) {
 	cfg := guardrailConfig{
-		ModelAllowlist: merged.ModelAllowlist,
-		PromptCapture: guardrailPromptCapture{
-			Enabled:   merged.PromptCapture.Enabled,
-			RedactPii: merged.PromptCapture.RedactPii,
-		},
+		ProviderAllowlists: providerAllowlists,
+		PromptCapture:      guardrailPromptCapture(capture),
 	}
 	out, err := json.Marshal(cfg)
 	if err != nil {
@@ -843,6 +855,74 @@ func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
 	return out, nil
 }

+// buildProviderAllowlists returns the proxy's per-provider backstop: a provider
+// is included only when every authorising policy restricts models (their union);
+// if any leaves it unrestricted it is omitted, so management decides per group.
+func buildProviderAllowlists(policies []*types.Policy, byID map[string]*types.Guardrail) map[string][]string {
+	type providerAcc struct {
+		models          map[string]struct{}
+		anyUnrestricted bool
+	}
+	accs := make(map[string]*providerAcc)
+	for _, p := range policies {
+		if p == nil {
+			continue
+		}
+		restricted, models := policyModelAllowlist(p, byID)
+		for _, providerID := range p.DestinationProviderIDs {
+			if providerID == "" {
+				continue
+			}
+			acc, ok := accs[providerID]
+			if !ok {
+				acc = &providerAcc{models: make(map[string]struct{})}
+				accs[providerID] = acc
+			}
+			if !restricted {
+				acc.anyUnrestricted = true
+				continue
+			}
+			for _, m := range models {
+				acc.models[m] = struct{}{}
+			}
+		}
+	}
+	out := make(map[string][]string, len(accs))
+	for providerID, acc := range accs {
+		if acc.anyUnrestricted {
+			continue
+		}
+		models := make([]string, 0, len(acc.models))
+		for m := range acc.models {
+			models = append(models, m)
+		}
+		sort.Strings(models)
+		out[providerID] = models
+	}
+	return out
+}
+
+// policyModelAllowlist reports whether a policy restricts models (has an
+// allowlist-enabled guardrail) and the union of allowed models. Models are
+// verbatim; the proxy factory lowercases/trims them at decode time.
+func policyModelAllowlist(p *types.Policy, byID map[string]*types.Guardrail) (bool, []string) {
+	restricted := false
+	var models []string
+	for _, gID := range p.GuardrailIDs {
+		g, ok := byID[gID]
+		if !ok || g == nil || !g.Checks.ModelAllowlist.Enabled {
+			continue
+		}
+		restricted = true
+		for _, m := range g.Checks.ModelAllowlist.Models {
+			if m != "" {
+				models = append(models, m)
+			}
+		}
+	}
+	return restricted, models
+}
+
 // buildAccountService composes the per-account gateway Service. The
 // target carries the noop placeholder URL — the router middleware
 // rewrites every request to the matched provider's upstream before the
@@ -986,38 +1066,11 @@ func unionSourceGroups(policies []*types.Policy) []string {
 	return out
 }

-// MergedGuardrails is the JSON shape passed to the proxy via the
-// guardrail middleware's config_json. Mirrors the proxy-side
-// expectations and is intentionally distinct from
-// types.GuardrailChecks so we can evolve either side independently.
+// MergedGuardrails is the synthesiser's fold target. Only prompt capture is
+// merged here — the model allowlist is emitted per-provider, and
+// token/budget/retention moved onto Policy.Limits and account Settings.
 type MergedGuardrails struct {
-	ModelAllowlist []string            `json:"model_allowlist,omitempty"`
-	TokenLimits    MergedTokenLimits   `json:"token_limits"`
-	Budget         MergedBudget        `json:"budget"`
-	PromptCapture  MergedPromptCapture `json:"prompt_capture"`
-	Retention      MergedRetention     `json:"retention"`
-}
-
-type MergedTokenLimits struct {
-	Hourly  *MergedTokenWindow `json:"hourly,omitempty"`
-	Daily   *MergedTokenWindow `json:"daily,omitempty"`
-	Monthly *MergedTokenWindow `json:"monthly,omitempty"`
-}
-
-type MergedTokenWindow struct {
-	MaxInputTokens  int `json:"max_input_tokens,omitempty"`
-	MaxOutputTokens int `json:"max_output_tokens,omitempty"`
-}
-
-type MergedBudget struct {
-	Hourly  *MergedBudgetWindow `json:"hourly,omitempty"`
-	Daily   *MergedBudgetWindow `json:"daily,omitempty"`
-	Monthly *MergedBudgetWindow `json:"monthly,omitempty"`
-}
-
-type MergedBudgetWindow struct {
-	SoftCapUSD float64 `json:"soft_cap_usd,omitempty"`
-	HardCapUSD float64 `json:"hard_cap_usd,omitempty"`
+	PromptCapture MergedPromptCapture
 }

 type MergedPromptCapture struct {
@@ -1025,64 +1078,31 @@ type MergedPromptCapture struct {
 	RedactPii bool `json:"redact_pii"`
 }

-type MergedRetention struct {
-	Enabled bool `json:"enabled"`
-	Days    int  `json:"days"`
-}
-
-// mergeGuardrails computes the effective guardrail spec applied at the
-// proxy, given the referencing policies and the account's guardrail
-// catalogue. Policy enabled-ness is the caller's responsibility — only
-// enabled policies should be passed in.
+// mergeGuardrails folds the referencing policies' guardrails into the
+// prompt-capture decision only. The model allowlist is enforced per-policy/group
+// in management and shipped per-provider; token/budget/retention live off
+// guardrails now.
 //
-// Merge rules:
-//   - Model allowlist:   union of allowlists across policies that enable it.
-//   - Token / Budget:    most-restrictive (min of non-zero caps) per window.
-//   - Prompt capture:    enabled if any policy enables it; redact_pii sticks
-//     if any enabling policy turns it on.
-//   - Retention:         enabled if any enables it; smallest non-zero days wins.
+// Merge rule — prompt capture: enabled if any policy enables it; redact_pii
+// sticks if any enabling policy turns it on.
 func mergeGuardrails(policies []*types.Policy, byID map[string]*types.Guardrail) MergedGuardrails {
 	merged := MergedGuardrails{}
-	allowlist := make(map[string]struct{})
-	allowlistEnabled := false
-
 	for _, policy := range policies {
 		for _, gID := range policy.GuardrailIDs {
 			g, ok := byID[gID]
 			if !ok || g == nil {
 				continue
 			}
-			mergeGuardrail(g, &merged, allowlist, &allowlistEnabled)
+			mergeGuardrail(g, &merged)
 		}
 	}
-
-	if allowlistEnabled {
-		merged.ModelAllowlist = make([]string, 0, len(allowlist))
-		for m := range allowlist {
-			merged.ModelAllowlist = append(merged.ModelAllowlist, m)
-		}
-		sort.Strings(merged.ModelAllowlist)
-	}
 	return merged
 }

-// mergeGuardrail folds a single guardrail's enabled checks into the
-// running merge: model-allowlist models join the shared set (and flip
-// allowlistEnabled), and prompt-capture / redact-pii stick once any
-// enabling guardrail turns them on.
-//
-// TokenLimits, Budget, and Retention have moved off guardrails — token
-// and budget caps now live on the Policy itself (Policy.Limits) and
-// retention moves to account-level Settings — so they are not merged here.
-func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails, allowlist map[string]struct{}, allowlistEnabled *bool) {
-	if g.Checks.ModelAllowlist.Enabled {
-		*allowlistEnabled = true
-		for _, m := range g.Checks.ModelAllowlist.Models {
-			if m != "" {
-				allowlist[m] = struct{}{}
-			}
-		}
-	}
+// mergeGuardrail folds a single guardrail's prompt-capture settings into the
+// running merge: enabled / redact-pii stick once any enabling guardrail turns
+// them on.
+func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails) {
 	if g.Checks.PromptCapture.Enabled {
 		merged.PromptCapture.Enabled = true
 		if g.Checks.PromptCapture.RedactPii {
@@ -81,8 +81,8 @@ func TestSynthesizeServices_RealStore_PromptCaptureAccountIsSoleControl(t *testi
 	require.Len(t, services, 1)

 	cfg := decodeServiceGuardrailConfig(t, services[0])
-	assert.Equal(t, []string{"gpt-5.4"}, cfg.ModelAllowlist,
-		"model allowlist is a pure policy guardrail and must always reach the config")
+	assert.Equal(t, map[string][]string{"prov-1": {"gpt-5.4"}}, cfg.ProviderAllowlists,
+		"model allowlist is a pure policy guardrail and must reach the per-provider config")
 	assert.False(t, cfg.PromptCapture.Enabled,
 		"prompt capture must be off when the account toggle is off, even with a capture-enabled guardrail")
 }
@@ -172,7 +172,7 @@ func TestSynthesizeServices_RealStore_NoGuardrail_CaptureOff(t *testing.T) {
 	require.Len(t, services, 1, "exactly one synth service expected")

 	cfg := decodeServiceGuardrailConfig(t, services[0])
-	assert.Empty(t, cfg.ModelAllowlist, "no guardrail → no allowlist")
+	assert.Empty(t, cfg.ProviderAllowlists, "no guardrail → provider unrestricted (absent from map)")
 	assert.False(t, cfg.PromptCapture.Enabled, "no guardrail → prompt capture off by default")
 	assert.False(t, cfg.PromptCapture.RedactPii, "no guardrail → redact off by default")
 }
@@ -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
+}
@@ -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")
+}
@@ -0,0 +1,95 @@
+package agentnetwork
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+)
+
+// policyForProviders builds an enabled policy authorising the given providers
+// under the given guardrails (both optional). Groups are irrelevant to
+// buildProviderAllowlists, which keys purely on destination provider.
+func policyForProviders(id string, guardrailIDs []string, providerIDs ...string) *types.Policy {
+	return &types.Policy{
+		ID:                     id,
+		Enabled:                true,
+		DestinationProviderIDs: providerIDs,
+		GuardrailIDs:           guardrailIDs,
+	}
+}
+
+func TestBuildProviderAllowlists(t *testing.T) {
+	byID := map[string]*types.Guardrail{
+		"g-4o":       allowlistGuardrail("g-4o", "acc-1", "gpt-4o"),
+		"g-opus":     allowlistGuardrail("g-opus", "acc-1", "claude-opus-4"),
+		"g-disabled": {ID: "g-disabled", Checks: types.GuardrailChecks{ModelAllowlist: types.GuardrailModelAllowlist{Enabled: false, Models: []string{"gpt-4o"}}}},
+	}
+
+	t.Run("all authorising policies restrict yields per-provider union", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", []string{"g-opus"}, "prov-x"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, map[string][]string{"prov-x": {"claude-opus-4", "gpt-4o"}}, got,
+			"a provider every policy restricts carries the sorted union of their models")
+	})
+
+	t.Run("any un-guardrailed policy leaves the provider unrestricted (omitted)", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", nil, "prov-x"), // no guardrail
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.NotContains(t, got, "prov-x",
+			"a provider reachable by an un-guardrailed policy must be omitted so the proxy treats it as unrestricted")
+	})
+
+	t.Run("a disabled allowlist counts as unrestricted", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-disabled"}, "prov-x"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.NotContains(t, got, "prov-x",
+			"a policy whose only guardrail has a disabled allowlist is unrestricted")
+	})
+
+	t.Run("providers are isolated from one another", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", []string{"g-opus"}, "prov-y"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-x"], "prov-x keeps only its own model")
+		assert.Equal(t, []string{"claude-opus-4"}, got["prov-y"], "prov-y keeps only its own model")
+	})
+
+	t.Run("one policy authorising two providers restricts both", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x", "prov-y"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-x"])
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-y"])
+	})
+
+	t.Run("union across a single policy's guardrails", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o", "g-opus"}, "prov-x"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.ElementsMatch(t, []string{"claude-opus-4", "gpt-4o"}, got["prov-x"],
+			"a policy's own multiple allowlist guardrails union together")
+	})
+
+	t.Run("an enabled allowlist with no models denies everything", func(t *testing.T) {
+		empty := map[string]*types.Guardrail{"g-empty": allowlistGuardrail("g-empty", "acc-1")}
+		got := buildProviderAllowlists([]*types.Policy{
+			policyForProviders("p1", []string{"g-empty"}, "prov-x"),
+		}, empty)
+		assert.Equal(t, map[string][]string{"prov-x": {}}, got,
+			"an enabled-but-empty allowlist is restricted with an empty set, not unrestricted")
+	})
+}
@@ -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")
@@ -1031,8 +1054,12 @@ func TestSynthesizeServices_GuardrailMerge_AllowlistUnion_LimitsRestrictive(t *t

 	var cfg guardrailConfig
 	require.NoError(t, json.Unmarshal(guardrailJSON, &cfg), "guardrail config must unmarshal cleanly")
-	assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ModelAllowlist,
-		"model allowlist union must keep both models")
+	// Both policies restrict the same provider, so the per-provider backstop
+	// carries the union of their models — a coarse gate that management's
+	// per-policy/group check narrows; it only blocks models outside the union
+	// when management is down.
+	assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ProviderAllowlists["prov-1"],
+		"per-provider allowlist union must keep both models")
 }

 func TestSynthesizeServices_BackfillsMissingSessionKeys(t *testing.T) {
@@ -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))
@@ -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")
@@ -36,7 +36,7 @@ func NewGRPCController(proxyGRPCServer *nbgrpc.ProxyServiceServer, meter metric.
 // SendServiceUpdateToCluster sends a service update to a specific proxy cluster.
 func (c *GRPCController) SendServiceUpdateToCluster(ctx context.Context, accountID string, update *proto.ProxyMapping, clusterAddr string) {
 	c.proxyGRPCServer.SendServiceUpdateToCluster(ctx, update, clusterAddr)
-	c.metrics.IncrementServiceUpdateSendCount(clusterAddr)
+	c.metrics.IncrementServiceUpdateSendCount()
 }

 // GetOIDCValidationConfig returns the OIDC validation configuration from the gRPC server.
@@ -53,7 +53,7 @@ func (c *GRPCController) RegisterProxyToCluster(ctx context.Context, clusterAddr
 	proxySet.(*sync.Map).Store(proxyID, struct{}{})
 	log.WithContext(ctx).Debugf("Registered proxy %s to cluster %s", proxyID, clusterAddr)

-	c.metrics.IncrementProxyConnectionCount(clusterAddr)
+	c.metrics.IncrementProxyConnectionCount()

 	return nil
 }
@@ -67,7 +67,7 @@ func (c *GRPCController) UnregisterProxyFromCluster(ctx context.Context, cluster
 		proxySet.(*sync.Map).Delete(proxyID)
 		log.WithContext(ctx).Debugf("Unregistered proxy %s from cluster %s", proxyID, clusterAddr)

-		c.metrics.DecrementProxyConnectionCount(clusterAddr)
+		c.metrics.DecrementProxyConnectionCount()
 	}
 	return nil
 }
@@ -3,7 +3,6 @@ package manager
 import (
 	"context"

-	"go.opentelemetry.io/otel/attribute"
 	"go.opentelemetry.io/otel/metric"
 )

@@ -48,25 +47,16 @@ func newMetrics(meter metric.Meter) (*metrics, error) {
 	}, nil
 }

-func (m *metrics) IncrementProxyConnectionCount(clusterAddr string) {
-	m.proxyConnectionCount.Add(context.Background(), 1,
-		metric.WithAttributes(
-			attribute.String("cluster", clusterAddr),
-		))
+func (m *metrics) IncrementProxyConnectionCount() {
+	m.proxyConnectionCount.Add(context.Background(), 1)
 }

-func (m *metrics) DecrementProxyConnectionCount(clusterAddr string) {
-	m.proxyConnectionCount.Add(context.Background(), -1,
-		metric.WithAttributes(
-			attribute.String("cluster", clusterAddr),
-		))
+func (m *metrics) DecrementProxyConnectionCount() {
+	m.proxyConnectionCount.Add(context.Background(), -1)
 }

-func (m *metrics) IncrementServiceUpdateSendCount(clusterAddr string) {
-	m.serviceUpdateSendCount.Add(context.Background(), 1,
-		metric.WithAttributes(
-			attribute.String("cluster", clusterAddr),
-		))
+func (m *metrics) IncrementServiceUpdateSendCount() {
+	m.serviceUpdateSendCount.Add(context.Background(), 1)
 }

 func (m *metrics) IncrementProxyHeartbeatCount() {
@@ -55,6 +55,8 @@ type Config struct {

 	ReverseProxy ReverseProxy

+	AgentNetwork AgentNetwork
+
 	// disable default all-to-all policy
 	DisableDefaultPolicy bool

@@ -185,6 +187,25 @@ 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). A relative path is
+	// resolved against <Datadir>, so a bare filename lands alongside the
+	// store. 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.
@@ -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()
@@ -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),
 	}
@@ -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")
+		})
+	}
+}
@@ -285,6 +285,7 @@ func (s *ProxyServiceServer) CheckLLMPolicyLimits(ctx context.Context, req *prot
 		UserID:     req.GetUserId(),
 		GroupIDs:   req.GetGroupIds(),
 		ProviderID: req.GetProviderId(),
+		Model:      req.GetModel(),
 	})
 	if err != nil {
 		log.WithContext(ctx).Errorf("select policy for request: %v", err)
@@ -0,0 +1,138 @@
+package grpc
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"google.golang.org/grpc/codes"
+	grpcstatus "google.golang.org/grpc/status"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
+	"github.com/netbirdio/netbird/shared/management/proto"
+)
+
+// fakeAgentNetworkLimits records the PolicySelectionInput it was invoked with
+// and returns a pre-programmed result, so tests can assert what the handler
+// forwards to the selector.
+type fakeAgentNetworkLimits struct {
+	gotInput agentnetwork.PolicySelectionInput
+	result   *agentnetwork.PolicySelectionResult
+	err      error
+}
+
+func (f *fakeAgentNetworkLimits) SelectPolicyForRequest(_ context.Context, in agentnetwork.PolicySelectionInput) (*agentnetwork.PolicySelectionResult, error) {
+	f.gotInput = in
+	if f.err != nil {
+		return nil, f.err
+	}
+	return f.result, nil
+}
+
+func (f *fakeAgentNetworkLimits) RecordUsage(_ context.Context, _ agentnetwork.RecordUsageInput) error {
+	return nil
+}
+
+// TestCheckLLMPolicyLimits_ForwardsModelToSelector proves the wiring added here:
+// the model the proxy extracted must reach the selector's Model unchanged,
+// alongside the account/user/group/provider fields.
+func TestCheckLLMPolicyLimits_ForwardsModelToSelector(t *testing.T) {
+	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{Allow: true, SelectedPolicyID: "pol-1"}}
+	s := &ProxyServiceServer{}
+	s.SetAgentNetworkLimitsService(fake)
+
+	req := &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		UserId:     "user-1",
+		GroupIds:   []string{"grp-a", "grp-b"},
+		ProviderId: "prov-1",
+		Model:      "claude-opus-4",
+	}
+
+	resp, err := s.CheckLLMPolicyLimits(context.Background(), req)
+	require.NoError(t, err)
+	require.NotNil(t, resp)
+
+	assert.Equal(t, "acc-1", fake.gotInput.AccountID)
+	assert.Equal(t, "user-1", fake.gotInput.UserID)
+	assert.Equal(t, []string{"grp-a", "grp-b"}, fake.gotInput.GroupIDs)
+	assert.Equal(t, "prov-1", fake.gotInput.ProviderID)
+	assert.Equal(t, "claude-opus-4", fake.gotInput.Model,
+		"the request's model must be forwarded to the selector")
+}
+
+// TestCheckLLMPolicyLimits_EmptyModelForwardedAsEmpty proves an undetermined
+// model (empty string) is forwarded as-is; the selector decides how to treat it.
+func TestCheckLLMPolicyLimits_EmptyModelForwardedAsEmpty(t *testing.T) {
+	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{Allow: true}}
+	s := &ProxyServiceServer{}
+	s.SetAgentNetworkLimitsService(fake)
+
+	req := &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		ProviderId: "prov-1",
+	}
+
+	_, err := s.CheckLLMPolicyLimits(context.Background(), req)
+	require.NoError(t, err)
+	assert.Equal(t, "", fake.gotInput.Model, "an absent model must be forwarded as empty, not fabricated")
+}
+
+// TestCheckLLMPolicyLimits_DenyResponseCarriesModelBlockedCode proves the deny
+// envelope surfaces the model-allowlist deny code + reason through the response.
+func TestCheckLLMPolicyLimits_DenyResponseCarriesModelBlockedCode(t *testing.T) {
+	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{
+		Allow:      false,
+		DenyCode:   "llm_policy.model_blocked",
+		DenyReason: `model "claude-opus-4" is not permitted by any applicable policy allowlist`,
+	}}
+	s := &ProxyServiceServer{}
+	s.SetAgentNetworkLimitsService(fake)
+
+	resp, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		ProviderId: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	require.NotNil(t, resp)
+	assert.Equal(t, "deny", resp.Decision)
+	assert.Equal(t, "llm_policy.model_blocked", resp.DenyCode)
+	assert.NotEmpty(t, resp.DenyReason)
+	assert.Empty(t, resp.SelectedPolicyId, "a denied request must carry no selected policy")
+}
+
+// TestCheckLLMPolicyLimits_SelectorErrorSurfacesAsInternal proves a selector
+// failure surfaces as an Internal gRPC error rather than a silent allow.
+func TestCheckLLMPolicyLimits_SelectorErrorSurfacesAsInternal(t *testing.T) {
+	fake := &fakeAgentNetworkLimits{err: errors.New("boom")}
+	s := &ProxyServiceServer{}
+	s.SetAgentNetworkLimitsService(fake)
+
+	_, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		ProviderId: "prov-1",
+		Model:      "gpt-4o",
+	})
+	require.Error(t, err)
+	st, ok := grpcstatus.FromError(err)
+	require.True(t, ok)
+	assert.Equal(t, codes.Internal, st.Code(), "selector errors must never fail open on the hot path")
+}
+
+// TestCheckLLMPolicyLimits_UnconfiguredServiceReturnsUnimplemented locks the
+// fallback: with no limits service wired the RPC returns Unimplemented.
+func TestCheckLLMPolicyLimits_UnconfiguredServiceReturnsUnimplemented(t *testing.T) {
+	s := &ProxyServiceServer{}
+
+	_, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		ProviderId: "prov-1",
+	})
+	require.Error(t, err)
+	st, ok := grpcstatus.FromError(err)
+	require.True(t, ok)
+	assert.Equal(t, codes.Unimplemented, st.Code())
+}
@@ -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),
 	}