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

2026-08-30 11:31:29 +02:00 · 2026-07-27 17:19:24 +09:00
parent 172914e5ee 4f6247b5c3
commit 82d31231e4
29 changed files with 1582 additions and 170 deletions
--- a/proxy/internal/accesslog/logger.go
+++ b/proxy/internal/accesslog/logger.go
@@ -221,14 +221,21 @@ func (l *Logger) allowDenyLog(serviceID types.ServiceID, reason string) bool {
 // proxy/internal/middleware/keys.go — only the dimensions management needs to
 // record a usage row (provider / model / tokens / cost / groups).
 var usageMetadataKeys = map[string]struct{}{
-	"llm.provider":             {},
-	"llm.model":                {},
-	"llm.resolved_provider_id": {},
-	"llm.input_tokens":         {},
-	"llm.output_tokens":        {},
-	"llm.total_tokens":         {},
-	"cost.usd_total":           {},
-	"llm.authorising_groups":   {},
+	"llm.provider":              {},
+	"llm.model":                 {},
+	"llm.resolved_provider_id":  {},
+	"llm.input_tokens":          {},
+	"llm.output_tokens":         {},
+	"llm.total_tokens":          {},
+	"llm.cached_input_tokens":   {},
+	"llm.cache_creation_tokens": {},
+	"cost.usd_input":            {},
+	"cost.usd_cached_input":     {},
+	"cost.usd_cache_creation":   {},
+	"cost.usd_output":           {},
+	"cost.usd_total":            {},
+	"cost.usd_cache":            {},
+	"llm.authorising_groups":    {},
 }

 // stripAgentNetworkEntryForUsage returns the entry reduced to what's needed to
--- a/proxy/internal/llm/bedrock.go
+++ b/proxy/internal/llm/bedrock.go
@@ -56,10 +56,12 @@ type bedrockResponse struct {
 		OutputTokens             int64 `json:"output_tokens"`
 		CacheReadInputTokens     int64 `json:"cache_read_input_tokens"`
 		CacheCreationInputTokens int64 `json:"cache_creation_input_tokens"`
-		// Converse — camelCase.
-		InputTokensCamel  int64 `json:"inputTokens"`
-		OutputTokensCamel int64 `json:"outputTokens"`
-		TotalTokensCamel  int64 `json:"totalTokens"`
+		// Converse — camelCase; cache buckets are additive to inputTokens (AWS names the write bucket cacheWriteInputTokens).
+		InputTokensCamel      int64 `json:"inputTokens"`
+		OutputTokensCamel     int64 `json:"outputTokens"`
+		TotalTokensCamel      int64 `json:"totalTokens"`
+		CacheReadTokensCamel  int64 `json:"cacheReadInputTokens"`
+		CacheWriteTokensCamel int64 `json:"cacheWriteInputTokens"`
 	} `json:"usage"`
 }

@@ -83,16 +85,18 @@ func (BedrockParser) ParseResponse(status int, contentType string, body []byte)
 	}
 	inTok := firstNonZero(resp.Usage.InputTokens, resp.Usage.InputTokensCamel)
 	outTok := firstNonZero(resp.Usage.OutputTokens, resp.Usage.OutputTokensCamel)
+	cacheRead := firstNonZero(resp.Usage.CacheReadInputTokens, resp.Usage.CacheReadTokensCamel)
+	cacheWrite := firstNonZero(resp.Usage.CacheCreationInputTokens, resp.Usage.CacheWriteTokensCamel)
 	total := resp.Usage.TotalTokensCamel
 	if total == 0 {
-		total = inTok + outTok + resp.Usage.CacheReadInputTokens + resp.Usage.CacheCreationInputTokens
+		total = inTok + outTok + cacheRead + cacheWrite
 	}
 	return Usage{
 		InputTokens:         inTok,
 		OutputTokens:        outTok,
 		TotalTokens:         total,
-		CachedInputTokens:   resp.Usage.CacheReadInputTokens,
-		CacheCreationTokens: resp.Usage.CacheCreationInputTokens,
+		CachedInputTokens:   cacheRead,
+		CacheCreationTokens: cacheWrite,
 	}, nil
 }

--- a/proxy/internal/llm/bedrock_test.go
+++ b/proxy/internal/llm/bedrock_test.go
@@ -26,6 +26,18 @@ func TestBedrockParser_ParseResponse_Converse(t *testing.T) {
 	require.Equal(t, int64(14), u.TotalTokens, "converse uses provider total")
 }

+// Converse camelCase cache fields must land in the billed Usage buckets, same as the InvokeModel snake_case fields.
+func TestBedrockParser_ParseResponse_ConverseCacheBuckets(t *testing.T) {
+	body := []byte(`{"usage":{"inputTokens":11,"outputTokens":3,"cacheReadInputTokens":7,"cacheWriteInputTokens":9}}`)
+	u, err := BedrockParser{}.ParseResponse(200, "application/json", body)
+	require.NoError(t, err)
+	require.Equal(t, int64(11), u.InputTokens, "converse input tokens")
+	require.Equal(t, int64(3), u.OutputTokens, "converse output tokens")
+	require.Equal(t, int64(7), u.CachedInputTokens, "converse cache-read tokens")
+	require.Equal(t, int64(9), u.CacheCreationTokens, "converse cache-write tokens")
+	require.Equal(t, int64(11+3+7+9), u.TotalTokens, "total backfill is additive when the provider omits totalTokens")
+}
+
 func TestBedrockParser_ParseResponse_StreamingUnsupported(t *testing.T) {
 	_, err := BedrockParser{}.ParseResponse(200, "application/vnd.amazon.eventstream", []byte("binary"))
 	require.ErrorIs(t, err, ErrStreamingUnsupported, "event-stream must route to the streaming accumulator")
--- a/proxy/internal/llm/pricing/pricing.go
+++ b/proxy/internal/llm/pricing/pricing.go
@@ -128,6 +128,46 @@ type Table struct {
 //   - Other providers: cached and cacheCreation are ignored; cost is
 //     inTokens*InputPer1K + outTokens*OutputPer1K.
 func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (float64, bool) {
+	c, ok := t.Costs(provider, model, inTokens, outTokens, cachedInput, cacheCreation)
+	return c.TotalUSD, ok
+}
+
+// Costs is a per-request cost split. The four per-bucket fields are the base
+// of the breakdown — one per token bucket the provider bills separately — and
+// the two aggregates are derived from them:
+//
+//	TotalUSD = InputUSD + CachedInputUSD + CacheCreationUSD + OutputUSD
+//	CacheUSD = CachedInputUSD + CacheCreationUSD
+//
+// InputUSD is always the cost of the *non-cached* input bucket, for both
+// provider shapes: on OpenAI the cached subset is carved out of inTokens and
+// billed as CachedInputUSD, so the two never double-count. Buckets a provider
+// doesn't bill are zero, which keeps the identities above true everywhere.
+type Costs struct {
+	InputUSD         float64
+	CachedInputUSD   float64
+	CacheCreationUSD float64
+	OutputUSD        float64
+	TotalUSD         float64
+	CacheUSD         float64
+}
+
+// newCosts assembles a split from its per-bucket parts, deriving the two
+// aggregates so TotalUSD and CacheUSD can never drift from the breakdown.
+func newCosts(input, cachedInput, cacheCreation, output float64) Costs {
+	return Costs{
+		InputUSD:         input,
+		CachedInputUSD:   cachedInput,
+		CacheCreationUSD: cacheCreation,
+		OutputUSD:        output,
+		TotalUSD:         input + cachedInput + cacheCreation + output,
+		CacheUSD:         cachedInput + cacheCreation,
+	}
+}
+
+// Costs returns the estimated USD cost split for the given token counts, with
+// the same semantics as Cost.
+func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (Costs, bool) {
 	// Clamp negatives to zero before any pricing math so a malformed
 	// upstream count can never produce a negative cost.
 	if inTokens < 0 {
@@ -143,15 +183,15 @@ func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, c
 		cacheCreation = 0
 	}
 	if t == nil {
-		return 0, false
+		return Costs{}, false
 	}
 	byModel, ok := t.entries[provider]
 	if !ok {
-		return 0, false
+		return Costs{}, false
 	}
 	entry, ok := byModel[model]
 	if !ok {
-		return 0, false
+		return Costs{}, false
 	}
 	output := (float64(outTokens) / 1000.0) * entry.OutputPer1K
 	switch provider {
@@ -168,7 +208,7 @@ func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, c
 		}
 		nonCached := float64(inTokens-clamped) / 1000.0 * entry.InputPer1K
 		cached := float64(clamped) / 1000.0 * cachedRate
-		return nonCached + cached + output, true
+		return newCosts(nonCached, cached, 0, output), true
 	case "anthropic", "bedrock":
 		// Bedrock-Anthropic returns the same additive cache buckets as
 		// first-party Anthropic; non-Anthropic Bedrock models simply report
@@ -184,10 +224,10 @@ func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, c
 		input := float64(inTokens) / 1000.0 * entry.InputPer1K
 		read := float64(cachedInput) / 1000.0 * readRate
 		create := float64(cacheCreation) / 1000.0 * createRate
-		return input + read + create + output, true
+		return newCosts(input, read, create, output), true
 	default:
 		input := float64(inTokens) / 1000.0 * entry.InputPer1K
-		return input + output, true
+		return newCosts(input, 0, 0, output), true
 	}
 }

--- a/proxy/internal/middleware/builtin/cost_calculation_matrix_test.go
+++ b/proxy/internal/middleware/builtin/cost_calculation_matrix_test.go
@@ -0,0 +1,329 @@
+package builtin_test
+
+import (
+	"bytes"
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"strconv"
+	"testing"
+
+	"github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/cost_meter"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_request_parser"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_response_parser"
+)
+
+// Drives the real pipeline (llm_request_parser → llm_response_parser → cost_meter) on the embedded default pricing
+// table and asserts exact USD amounts hardcoded from the vendors' published prices, including the cache split.
+func TestCostCalculation_ProviderMatrix(t *testing.T) {
+	// Empty data dir → embedded defaults, like a proxy with no pricing override.
+	builtin.Configure(context.Background(), t.TempDir(), nil, nil, nil)
+
+	reqMW, err := llm_request_parser.Factory{}.New(nil)
+	require.NoError(t, err, "build llm_request_parser")
+	respMW, err := llm_response_parser.Factory{}.New(nil)
+	require.NoError(t, err, "build llm_response_parser")
+	costMW, err := cost_meter.Factory{}.New(nil)
+	require.NoError(t, err, "build cost_meter")
+	t.Cleanup(func() { _ = costMW.Close() })
+
+	const jsonCT = "application/json"
+	const sseCT = "text/event-stream"
+	const awsCT = "application/vnd.amazon.eventstream"
+
+	cases := []struct {
+		name     string
+		url      string
+		reqBody  []byte
+		respCT   string
+		respBody []byte
+
+		wantProvider  string
+		wantModel     string
+		wantCost      float64 // exact expected USD; ignored when wantSkip is set
+		wantCacheCost float64 // expected cost.usd_cache portion of wantCost
+		wantSkip      string  // expected cost.skipped reason, "" when priced
+	}{
+		{
+			// gpt-4o-mini $0.15/$0.60 per MTok: 1000×0.15/1M + 500×0.60/1M.
+			name:         "openai chat completions",
+			url:          "https://api.openai.com/v1/chat/completions",
+			reqBody:      []byte(`{"model":"gpt-4o-mini","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"choices":[{"message":{"content":"pong"}}],"usage":{"prompt_tokens":1000,"completion_tokens":500,"total_tokens":1500}}`),
+			wantProvider: "openai",
+			wantModel:    "gpt-4o-mini",
+			wantCost:     0.00045,
+		},
+		{
+			// OpenAI cached tokens are a SUBSET of prompt_tokens at a discount; gpt-4o $2.50/$10 per MTok, cached $1.25/M:
+			// 250×2.5/1M + 750×1.25/1M + 500×10/1M.
+			name:          "openai cached subset discount",
+			url:           "https://api.openai.com/v1/chat/completions",
+			reqBody:       []byte(`{"model":"gpt-4o","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:        jsonCT,
+			respBody:      []byte(`{"usage":{"prompt_tokens":1000,"completion_tokens":500,"prompt_tokens_details":{"cached_tokens":750}}}`),
+			wantProvider:  "openai",
+			wantModel:     "gpt-4o",
+			wantCost:      0.0065625,
+			wantCacheCost: 0.0009375,
+		},
+		{
+			// OpenAI streaming: usage rides the final SSE frame.
+			name:         "openai chat SSE stream",
+			url:          "https://api.openai.com/v1/chat/completions",
+			reqBody:      []byte(`{"model":"gpt-4o-mini","stream":true,"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       sseCT,
+			respBody:     sseBody(`{"choices":[{"delta":{"content":"po"}}]}`, `{"choices":[{"delta":{"content":"ng"}}]}`, `{"choices":[],"usage":{"prompt_tokens":1000,"completion_tokens":500}}`, "[DONE]"),
+			wantProvider: "openai",
+			wantModel:    "gpt-4o-mini",
+			wantCost:     0.00045,
+		},
+		{
+			// Mistral speaks the OpenAI shape: mistral-large-latest $0.50/$1.50 per MTok.
+			name:         "mistral via openai shape",
+			url:          "https://api.mistral.ai/v1/chat/completions",
+			reqBody:      []byte(`{"model":"mistral-large-latest","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"usage":{"prompt_tokens":1000,"completion_tokens":1000}}`),
+			wantProvider: "openai",
+			wantModel:    "mistral-large-latest",
+			wantCost:     0.002,
+		},
+		{
+			// The field report, minus caching: Bedrock Sonnet 4.6 $3/$15 per MTok, 3×3/1M + 1514×15/1M = $0.022719.
+			// Also covers inference-profile normalization of the region-prefixed versioned id in the URL.
+			name:         "bedrock invoke — reported scenario, no cache",
+			url:          "https://bedrock-runtime.eu-central-1.amazonaws.com/model/global.anthropic.claude-sonnet-4-6-20260115-v1:0/invoke",
+			reqBody:      []byte(`{"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"content":[{"type":"text","text":"pong"}],"usage":{"input_tokens":3,"output_tokens":1514}}`),
+			wantProvider: "bedrock",
+			wantModel:    "anthropic.claude-sonnet-4-6",
+			wantCost:     0.022719,
+		},
+		{
+			// The field report as observed: the FIRST call of a session also wrote a 30,528-token prompt cache at
+			// 1.25× input ($3.75/M): 0.022719 + 30528×3.75/1M = $0.137199 — the reported $0.1372.
+			name:          "bedrock invoke — reported scenario with cache write",
+			url:           "https://bedrock-runtime.eu-central-1.amazonaws.com/model/global.anthropic.claude-sonnet-4-6-20260115-v1:0/invoke",
+			reqBody:       []byte(`{"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:        jsonCT,
+			respBody:      []byte(`{"usage":{"input_tokens":3,"output_tokens":1514,"cache_creation_input_tokens":30528,"cache_read_input_tokens":0}}`),
+			wantProvider:  "bedrock",
+			wantModel:     "anthropic.claude-sonnet-4-6",
+			wantCost:      0.137199,
+			wantCacheCost: 0.11448,
+		},
+		{
+			// Same numbers over the InvokeModel event-stream: message_start carries input + cache, message_delta the output.
+			name:          "bedrock invoke stream with cache write",
+			url:           "https://bedrock-runtime.eu-central-1.amazonaws.com/model/global.anthropic.claude-sonnet-4-6-20260115-v1:0/invoke-with-response-stream",
+			reqBody:       []byte(`{"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:        awsCT,
+			respBody:      bedrockInvokeStream(t, `{"type":"message_start","message":{"usage":{"input_tokens":3,"output_tokens":1,"cache_creation_input_tokens":30528}}}`, `{"type":"content_block_delta","delta":{"type":"text_delta","text":"pong"}}`, `{"type":"message_delta","usage":{"output_tokens":1514}}`),
+			wantProvider:  "bedrock",
+			wantModel:     "anthropic.claude-sonnet-4-6",
+			wantCost:      0.137199,
+			wantCacheCost: 0.11448,
+		},
+		{
+			// Converse camelCase usage incl. cache buckets. Haiku 4.5 $1/$5 per MTok, read $0.10/M, write $1.25/M:
+			// 50×1/1M + 100×5/1M + 2000×0.1/1M + 1000×1.25/1M = $0.002.
+			name:          "bedrock converse with cache buckets",
+			url:           "https://bedrock-runtime.eu-central-1.amazonaws.com/model/eu.anthropic.claude-haiku-4-5-20251001-v1:0/converse",
+			reqBody:       []byte(`{"messages":[{"role":"user","content":[{"text":"hi"}]}]}`),
+			respCT:        jsonCT,
+			respBody:      []byte(`{"output":{"message":{"content":[{"text":"pong"}]}},"usage":{"inputTokens":50,"outputTokens":100,"totalTokens":3150,"cacheReadInputTokens":2000,"cacheWriteInputTokens":1000}}`),
+			wantProvider:  "bedrock",
+			wantModel:     "anthropic.claude-haiku-4-5",
+			wantCost:      0.002,
+			wantCacheCost: 0.00145,
+		},
+		{
+			// Same numbers over converse-stream: usage rides the trailing metadata frame.
+			name:    "bedrock converse stream with cache buckets",
+			url:     "https://bedrock-runtime.eu-central-1.amazonaws.com/model/eu.anthropic.claude-haiku-4-5-20251001-v1:0/converse-stream",
+			reqBody: []byte(`{"messages":[{"role":"user","content":[{"text":"hi"}]}]}`),
+			respCT:  awsCT,
+			respBody: bedrockConverseStream(t,
+				`{"delta":{"text":"pong"}}`,
+				`{"usage":{"inputTokens":50,"outputTokens":100,"totalTokens":3150,"cacheReadInputTokens":2000,"cacheWriteInputTokens":1000}}`,
+			),
+			wantProvider:  "bedrock",
+			wantModel:     "anthropic.claude-haiku-4-5",
+			wantCost:      0.002,
+			wantCacheCost: 0.00145,
+		},
+		{
+			// First-party Anthropic, additive cache buckets. Sonnet 4.6:
+			// 256×3/1M + 200×15/1M + 768×0.3/1M + 512×3.75/1M.
+			name:          "anthropic messages with cache buckets",
+			url:           "https://api.anthropic.com/v1/messages",
+			reqBody:       []byte(`{"model":"claude-sonnet-4-6","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:        jsonCT,
+			respBody:      []byte(`{"content":[{"type":"text","text":"pong"}],"usage":{"input_tokens":256,"output_tokens":200,"cache_read_input_tokens":768,"cache_creation_input_tokens":512}}`),
+			wantProvider:  "anthropic",
+			wantModel:     "claude-sonnet-4-6",
+			wantCost:      0.0059184,
+			wantCacheCost: 0.0021504,
+		},
+		{
+			// Anthropic SSE: input from message_start, output from message_delta. Haiku 4.5: 1000×1/1M + 2000×5/1M.
+			name:         "anthropic SSE stream",
+			url:          "https://api.anthropic.com/v1/messages",
+			reqBody:      []byte(`{"model":"claude-haiku-4-5","stream":true,"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       sseCT,
+			respBody:     sseBody(`{"type":"message_start","message":{"usage":{"input_tokens":1000,"output_tokens":2}}}`, `{"type":"content_block_delta","delta":{"type":"text_delta","text":"pong"}}`, `{"type":"message_delta","usage":{"output_tokens":2000}}`, `{"type":"message_stop"}`),
+			wantProvider: "anthropic",
+			wantModel:    "claude-haiku-4-5",
+			wantCost:     0.011,
+		},
+		{
+			// Kimi's Anthropic-compatible endpoint: kimi-k3 $3/$15 per MTok under the anthropic table.
+			name:         "kimi anthropic shape",
+			url:          "https://api.moonshot.ai/anthropic/v1/messages",
+			reqBody:      []byte(`{"model":"kimi-k3","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"usage":{"input_tokens":1000,"output_tokens":1000}}`),
+			wantProvider: "anthropic",
+			wantModel:    "kimi-k3",
+			wantCost:     0.018,
+		},
+		{
+			// Vertex path-routed model with "@version" stripped; Anthropic-on-Vertex priced under the anthropic table.
+			name:         "vertex anthropic path-routed",
+			url:          "https://aiplatform.googleapis.com/v1/projects/p/locations/global/publishers/anthropic/models/claude-sonnet-4-6@20260115:rawPredict",
+			reqBody:      []byte(`{"anthropic_version":"vertex-2023-10-16","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"usage":{"input_tokens":200,"output_tokens":100}}`),
+			wantProvider: "anthropic",
+			wantModel:    "claude-sonnet-4-6",
+			wantCost:     0.0021,
+		},
+		{
+			// Gateway-prefixed model ids are not in the pricing table: the meter must SKIP, never guess a rate.
+			name:         "gateway-prefixed model skips pricing",
+			url:          "https://gateway.example.com/v1/chat/completions",
+			reqBody:      []byte(`{"model":"openai/gpt-4o-mini","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"usage":{"prompt_tokens":1000,"completion_tokens":500}}`),
+			wantProvider: "openai",
+			wantModel:    "openai/gpt-4o-mini",
+			wantSkip:     "unknown_model",
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			in := &middleware.Input{
+				Method:  "POST",
+				URL:     tc.url,
+				Headers: []middleware.KV{{Key: "Content-Type", Value: "application/json"}},
+				Body:    tc.reqBody,
+			}
+
+			reqOut, err := reqMW.Invoke(context.Background(), in)
+			require.NoError(t, err, "request parser")
+			in.Metadata = append(in.Metadata, reqOut.Metadata...)
+
+			require.Equal(t, tc.wantProvider, metaKV(in.Metadata, middleware.KeyLLMProvider), "detected provider")
+			require.Equal(t, tc.wantModel, metaKV(in.Metadata, middleware.KeyLLMModel), "detected (normalized) model")
+
+			in.Status = 200
+			in.RespHeaders = []middleware.KV{{Key: "Content-Type", Value: tc.respCT}}
+			in.RespBody = tc.respBody
+
+			respOut, err := respMW.Invoke(context.Background(), in)
+			require.NoError(t, err, "response parser")
+			in.Metadata = append(in.Metadata, respOut.Metadata...)
+
+			costOut, err := costMW.Invoke(context.Background(), in)
+			require.NoError(t, err, "cost meter")
+
+			if tc.wantSkip != "" {
+				assert.Equal(t, tc.wantSkip, metaKV(costOut.Metadata, middleware.KeyCostSkipped), "expected cost skip reason")
+				assert.Empty(t, metaKV(costOut.Metadata, middleware.KeyCostUSDTotal), "no cost may be emitted on skip")
+				return
+			}
+
+			raw := metaKV(costOut.Metadata, middleware.KeyCostUSDTotal)
+			require.NotEmpty(t, raw, "cost.usd_total must be emitted; skip=%q", metaKV(costOut.Metadata, middleware.KeyCostSkipped))
+			got, err := strconv.ParseFloat(raw, 64)
+			require.NoError(t, err, "cost must be a float")
+			// cost.usd_total is rendered with %.6f: allow half of the last printed digit on top of float error.
+			assert.InDelta(t, tc.wantCost, got, 5.1e-7, "USD cost for %s", tc.name)
+
+			rawCache := metaKV(costOut.Metadata, middleware.KeyCostUSDCache)
+			require.NotEmpty(t, rawCache, "cost.usd_cache must be emitted next to cost.usd_total")
+			gotCache, err := strconv.ParseFloat(rawCache, 64)
+			require.NoError(t, err, "cache cost must be a float")
+			assert.InDelta(t, tc.wantCacheCost, gotCache, 5.1e-7, "cache USD cost for %s", tc.name)
+		})
+	}
+}
+
+// metaKV returns the value for key in kvs, or "" when absent.
+func metaKV(kvs []middleware.KV, key string) string {
+	for _, kv := range kvs {
+		if kv.Key == key {
+			return kv.Value
+		}
+	}
+	return ""
+}
+
+// sseBody renders data frames as a text/event-stream body.
+func sseBody(frames ...string) []byte {
+	var b bytes.Buffer
+	for _, f := range frames {
+		b.WriteString("data: ")
+		b.WriteString(f)
+		b.WriteString("\n\n")
+	}
+	return b.Bytes()
+}
+
+// awsFrame encodes one AWS event-stream frame with the given :event-type.
+func awsFrame(t *testing.T, eventType string, payload []byte) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	enc := eventstream.NewEncoder()
+	require.NoError(t, enc.Encode(&buf, eventstream.Message{
+		Headers: eventstream.Headers{{Name: ":event-type", Value: eventstream.StringValue(eventType)}},
+		Payload: payload,
+	}), "encode event-stream frame")
+	return buf.Bytes()
+}
+
+// bedrockInvokeStream builds an invoke-with-response-stream body: each "chunk" frame wraps a base64 Anthropic event.
+func bedrockInvokeStream(t *testing.T, events ...string) []byte {
+	t.Helper()
+	var body bytes.Buffer
+	for _, ev := range events {
+		wrap, err := json.Marshal(map[string]string{"bytes": base64.StdEncoding.EncodeToString([]byte(ev))})
+		require.NoError(t, err)
+		body.Write(awsFrame(t, "chunk", wrap))
+	}
+	return body.Bytes()
+}
+
+// bedrockConverseStream builds a converse-stream body: contentBlockDelta frames plus a trailing metadata usage frame.
+func bedrockConverseStream(t *testing.T, deltas ...string) []byte {
+	t.Helper()
+	var body bytes.Buffer
+	for i, ev := range deltas {
+		eventType := "contentBlockDelta"
+		if i == len(deltas)-1 {
+			eventType = "metadata"
+		}
+		body.Write(awsFrame(t, eventType, []byte(ev)))
+	}
+	return body.Bytes()
+}
--- a/proxy/internal/middleware/builtin/cost_meter/middleware.go
+++ b/proxy/internal/middleware/builtin/cost_meter/middleware.go
@@ -32,7 +32,12 @@ const (
 )

 var metadataKeys = []string{
+	middleware.KeyCostUSDInput,
+	middleware.KeyCostUSDCachedInput,
+	middleware.KeyCostUSDCacheCreation,
+	middleware.KeyCostUSDOutput,
 	middleware.KeyCostUSDTotal,
+	middleware.KeyCostUSDCache,
 	middleware.KeyCostSkipped,
 }

@@ -140,18 +145,38 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 	}

 	table := m.loader.Get()
-	cost, ok := table.Cost(provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
+	costs, ok := table.Costs(provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
 	if !ok {
 		out.Metadata = skip(skipUnknownModel)
 		return out, nil
 	}

+	// Per-bucket costs first: they're the base of the breakdown, and the two
+	// aggregates that follow are derived from exactly these four values.
 	out.Metadata = []middleware.KV{
-		{Key: middleware.KeyCostUSDTotal, Value: fmt.Sprintf("%.6f", cost)},
+		{Key: middleware.KeyCostUSDInput, Value: usd(costs.InputUSD)},
+		{Key: middleware.KeyCostUSDCachedInput, Value: usd(costs.CachedInputUSD)},
+		{Key: middleware.KeyCostUSDCacheCreation, Value: usd(costs.CacheCreationUSD)},
+		{Key: middleware.KeyCostUSDOutput, Value: usd(costs.OutputUSD)},
+		{Key: middleware.KeyCostUSDTotal, Value: usd(costs.TotalUSD)},
+		{Key: middleware.KeyCostUSDCache, Value: usd(costs.CacheUSD)},
 	}
 	return out, nil
 }

+// usd renders a cost as the fixed-precision string every cost.usd_* key
+// carries, so the per-bucket values and the aggregates round identically.
+//
+// 9 decimals, not 6: these values are summed downstream — per request, per
+// session, and per usage bucket — so the rounding step is applied once per
+// bucket per row and then accumulated. At 6 decimals a single row loses up to
+// 2e-6 across its four buckets (enough to break a 1e-6 reconciliation against
+// published rates), and a bucket smaller than half a microdollar quantises to
+// zero outright: 16 cache-read tokens on a cheap model is 1.6e-9, so summing
+// 10k such rows reports 0.02 instead of 0.016. Nano-dollar precision keeps the
+// per-row error ~1000x below the smallest realistic bucket.
+func usd(v float64) string { return fmt.Sprintf("%.9f", v) }
+
 // skip returns a single-entry metadata slice carrying the given skip
 // reason under KeyCostSkipped.
 func skip(reason string) []middleware.KV {
--- a/proxy/internal/middleware/builtin/cost_meter/middleware_test.go
+++ b/proxy/internal/middleware/builtin/cost_meter/middleware_test.go
@@ -67,7 +67,15 @@ func TestMiddleware_StaticSurface(t *testing.T) {
 	assert.NoError(t, mw.Close(), "Close on stateless middleware is a no-op")

 	keys := mw.MetadataKeys()
-	expected := []string{middleware.KeyCostUSDTotal, middleware.KeyCostSkipped}
+	expected := []string{
+		middleware.KeyCostUSDInput,
+		middleware.KeyCostUSDCachedInput,
+		middleware.KeyCostUSDCacheCreation,
+		middleware.KeyCostUSDOutput,
+		middleware.KeyCostUSDTotal,
+		middleware.KeyCostUSDCache,
+		middleware.KeyCostSkipped,
+	}
 	assert.Equal(t, expected, keys, "metadata key allowlist must match the spec")
 }

@@ -105,7 +113,7 @@ func TestFactory_DefaultPricingPathLoadsFixture(t *testing.T) {

 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "cost.usd_total must be emitted for known model")
-	assert.Equal(t, "0.000750", value, "0.00015 + 0.0006 per 1k tokens, 6-decimal format")
+	assert.Equal(t, "0.000750000", value, "0.00015 + 0.0006 per 1k tokens, 9-decimal format")
 }

 func TestFactory_PricingPathOverride(t *testing.T) {
@@ -129,7 +137,7 @@ func TestFactory_PricingPathOverride(t *testing.T) {

 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "cost.usd_total must be emitted with custom pricing path")
-	assert.Equal(t, "0.015000", value, "2*0.0025 + 1*0.01 = 0.015 with 6-decimal format")
+	assert.Equal(t, "0.015000000", value, "2*0.0025 + 1*0.01 = 0.015 with 9-decimal format")
 }

 func TestInvoke_ComputesCostForKnownModel(t *testing.T) {
@@ -148,7 +156,7 @@ func TestInvoke_ComputesCostForKnownModel(t *testing.T) {

 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "cost.usd_total must be emitted")
-	assert.Equal(t, "0.018000", value, "0.003 + 0.015 = 0.018 with 6-decimal format")
+	assert.Equal(t, "0.018000000", value, "0.003 + 0.015 = 0.018 with 9-decimal format")
 	_, skipped := metaValue(t, out.Metadata, middleware.KeyCostSkipped)
 	assert.False(t, skipped, "cost.skipped must not be set when cost is computed")
 }
@@ -357,8 +365,25 @@ func TestInvoke_OpenAICachedSubsetDiscount(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "cached subset path must produce a cost — never a skip")
 	// 250 non-cached at 0.0025/1k + 750 cached at 0.00125/1k + 500 output at 0.01/1k.
-	assert.Equal(t, "0.006563", value,
+	assert.Equal(t, "0.006562500", value,
 		"cached subset must be billed at the discount rate, non-cached at the full rate; never double-billed")
+
+	cache, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDCache)
+	require.True(t, ok, "cost.usd_cache must be emitted alongside cost.usd_total")
+	// 750 cached at 0.00125/1k = 0.0009375.
+	assert.Equal(t, "0.000937500", cache, "cache cost is the discounted cost of the cached subset")
+
+	// Per-bucket breakdown. On OpenAI the cached subset is carved out of the
+	// input bucket, so input covers only the 250 non-cached tokens — the two
+	// must never double-count the same 750 tokens.
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDInput, "0.000625000",
+		"input bucket bills only the non-cached remainder")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDCachedInput, "0.000937500",
+		"cached-input bucket bills the discounted subset")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDCacheCreation, "0.000000000",
+		"OpenAI has no cache-write bucket")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDOutput, "0.005000000",
+		"output bucket bills 500 tokens at 0.01/1k")
 }

 // TestInvoke_AnthropicCacheBucketsAdditive proves the Anthropic
@@ -384,9 +409,33 @@ func TestInvoke_AnthropicCacheBucketsAdditive(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok)
 	// 256 input * 0.003 + 768 cache_read * 0.0003 + 512 cache_creation * 0.00375 + 200 output * 0.015
-	// = 0.000768 + 0.0002304 + 0.00192 + 0.003 = 0.0059184 → "0.005918" with 6-decimal format.
-	assert.Equal(t, "0.005918", value,
+	// = 0.000768 + 0.0002304 + 0.00192 + 0.003 = 0.0059184.
+	assert.Equal(t, "0.005918400", value,
 		"each Anthropic input bucket must bill at its own rate — cache_read cheap, cache_creation expensive, regular input mid")
+
+	cache, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDCache)
+	require.True(t, ok, "cost.usd_cache must be emitted alongside cost.usd_total")
+	// 768 cache_read * 0.0003 + 512 cache_creation * 0.00375 = 0.0021504.
+	assert.Equal(t, "0.002150400", cache, "cache cost sums the read and creation buckets")
+
+	// Per-bucket breakdown: four separately-billed buckets, each at its own rate.
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDInput, "0.000768000",
+		"input bucket bills 256 tokens at 0.003/1k")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDCachedInput, "0.000230400",
+		"cache-read bucket bills 768 tokens at the cheap 0.0003/1k")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDCacheCreation, "0.001920000",
+		"cache-write bucket bills 512 tokens at the expensive 0.00375/1k")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDOutput, "0.003000000",
+		"output bucket bills 200 tokens at 0.015/1k")
+}
+
+// assertBucket asserts one per-bucket cost key carries the expected
+// 6-decimal value.
+func assertBucket(t *testing.T, md []middleware.KV, key, want, msg string) {
+	t.Helper()
+	got, ok := metaValue(t, md, key)
+	require.Truef(t, ok, "%s must be emitted", key)
+	assert.Equal(t, want, got, msg)
 }

 // TestInvoke_CachedTokensAbsentFallsBackToBaseFormula covers the
@@ -411,7 +460,7 @@ func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok)
 	// 1000 input * 0.0025 + 500 output * 0.01 = 0.0025 + 0.005 = 0.0075
-	assert.Equal(t, "0.007500", value, "no cached metadata = same cost as before the feature landed")
+	assert.Equal(t, "0.007500000", value, "no cached metadata = same cost as before the feature landed")
 }

 // TestInvoke_UnparseableCachedTokensSkippedSilently proves the
@@ -435,7 +484,7 @@ func TestInvoke_UnparseableCachedTokensSkippedSilently(t *testing.T) {
 	require.NoError(t, err)
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "garbage cache metadata must NOT switch the response from a cost to a skip — fall back to 0 cached")
-	assert.Equal(t, "0.007500", value, "same as the no-cached-metadata path")
+	assert.Equal(t, "0.007500000", value, "same as the no-cached-metadata path")
 }

 // TestMiddleware_CloseCancelsReloader proves Close stops the per-instance
--- a/proxy/internal/middleware/builtin/llm_response_parser/streaming_bedrock.go
+++ b/proxy/internal/middleware/builtin/llm_response_parser/streaming_bedrock.go
@@ -69,15 +69,18 @@ func applyBedrockInvokeChunk(payload []byte, usage *llm.Usage, completion *strin
 }

 // converseStreamEvent captures the Converse stream frames carrying completion
-// text (contentBlockDelta) and the final token usage (metadata).
+// text (contentBlockDelta) and the final token usage (metadata). Cache buckets
+// are additive to inputTokens (AWS write bucket: cacheWriteInputTokens).
 type converseStreamEvent struct {
 	Delta *struct {
 		Text string `json:"text"`
 	} `json:"delta"`
 	Usage *struct {
-		InputTokens  int64 `json:"inputTokens"`
-		OutputTokens int64 `json:"outputTokens"`
-		TotalTokens  int64 `json:"totalTokens"`
+		InputTokens      int64 `json:"inputTokens"`
+		OutputTokens     int64 `json:"outputTokens"`
+		TotalTokens      int64 `json:"totalTokens"`
+		CacheReadTokens  int64 `json:"cacheReadInputTokens"`
+		CacheWriteTokens int64 `json:"cacheWriteInputTokens"`
 	} `json:"usage"`
 }

@@ -105,6 +108,12 @@ func applyConverseStreamEvent(eventType string, payload []byte, usage *llm.Usage
 			if ev.Usage.TotalTokens > 0 {
 				usage.TotalTokens = ev.Usage.TotalTokens
 			}
+			if ev.Usage.CacheReadTokens > 0 {
+				usage.CachedInputTokens = ev.Usage.CacheReadTokens
+			}
+			if ev.Usage.CacheWriteTokens > 0 {
+				usage.CacheCreationTokens = ev.Usage.CacheWriteTokens
+			}
 		}
 	}
 }
--- a/proxy/internal/middleware/builtin/llm_response_parser/streaming_bedrock_test.go
+++ b/proxy/internal/middleware/builtin/llm_response_parser/streaming_bedrock_test.go
@@ -66,6 +66,24 @@ func TestAccumulateBedrockStream_Converse(t *testing.T) {
 	require.Equal(t, "pong", completion, "converse text deltas concatenated")
 }

+// The converse-stream metadata frame's camelCase cache fields must reach the billed cache buckets.
+func TestAccumulateBedrockStream_ConverseCacheBuckets(t *testing.T) {
+	var body bytes.Buffer
+	body.Write(bedrockFrame(t, "contentBlockDelta", mustJSON(t, map[string]any{"delta": map[string]any{"text": "pong"}})))
+	body.Write(bedrockFrame(t, "metadata", mustJSON(t, map[string]any{"usage": map[string]any{
+		"inputTokens": 11, "outputTokens": 3, "totalTokens": 30,
+		"cacheReadInputTokens": 7, "cacheWriteInputTokens": 9,
+	}})))
+
+	usage, completion := accumulateBedrockStream(body.Bytes())
+	require.Equal(t, int64(11), usage.InputTokens, "input tokens from metadata frame")
+	require.Equal(t, int64(3), usage.OutputTokens, "output tokens from metadata frame")
+	require.Equal(t, int64(7), usage.CachedInputTokens, "cache-read tokens from metadata frame")
+	require.Equal(t, int64(9), usage.CacheCreationTokens, "cache-write tokens from metadata frame")
+	require.Equal(t, int64(30), usage.TotalTokens, "provider-reported total wins")
+	require.Equal(t, "pong", completion)
+}
+
 func TestAccumulateBedrockStream_Truncated(t *testing.T) {
 	// A body cut mid-frame must not panic; partial usage is returned.
 	full := bedrockFrame(t, "metadata", mustJSON(t, map[string]any{"usage": map[string]any{"inputTokens": 11, "outputTokens": 3}}))
--- a/proxy/internal/middleware/keys.go
+++ b/proxy/internal/middleware/keys.go
@@ -75,8 +75,19 @@ const (
 	KeyLLMAttributionGroupID = "llm.attribution_group_id"
 	KeyLLMAttributionWindowS = "llm.attribution_window_seconds"

-	// Cost metering (emitted by cost_meter).
-	KeyCostUSDTotal = "cost.usd_total"
+	// Cost metering (emitted by cost_meter). The four per-bucket keys are the
+	// base of the breakdown — one per token bucket the provider bills
+	// separately — and the two aggregates below are derived from them:
+	// usd_total is their sum, usd_cache is cached_input + cache_creation.
+	KeyCostUSDInput = "cost.usd_input"
+	// KeyCostUSDCachedInput is the cost of the cache-read bucket (Anthropic cache_read; OpenAI's discounted cached subset of input).
+	KeyCostUSDCachedInput = "cost.usd_cached_input"
+	// KeyCostUSDCacheCreation is the cost of the cache-write bucket. Zero for providers without one.
+	KeyCostUSDCacheCreation = "cost.usd_cache_creation"
+	KeyCostUSDOutput        = "cost.usd_output"
+	KeyCostUSDTotal         = "cost.usd_total"
+	// KeyCostUSDCache is the portion of cost.usd_total billed for prompt-cache buckets (cache read/creation, or OpenAI's cached input subset).
+	KeyCostUSDCache = "cost.usd_cache"
 	KeyCostSkipped  = "cost.skipped"

 	// Framework-emitted error markers. Use the mw.<id>.* prefix to