From c906bf000e0c425dcc1fa63f4f7989f92d9ba4d5 Mon Sep 17 00:00:00 2001
From: alfadb <alfadb@163.com>
Date: Sat, 2 May 2026 11:14:25 +0800
Subject: [PATCH 1/8] feat(billing): add DeepSeek V4 Pro / Flash fallback
 pricing
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Pricing source: https://api-docs.deepseek.com/quick_start/pricing

deepseek-v4-pro:  input $0.435/M, output $0.87/M, cache_read $0.003625/M
deepseek-v4-flash: input $0.14/M,  output $0.28/M, cache_read $0.0028/M

Also map legacy aliases deepseek-chat / deepseek-reasoner → v4-flash.
---
 backend/internal/service/billing_service.go   | 30 +++++++++++++++++++
 .../internal/service/billing_service_test.go  |  4 +++
 2 files changed, 34 insertions(+)

diff --git a/backend/internal/service/billing_service.go b/backend/internal/service/billing_service.go
index 75771f23b9..c4e376fcff 100644
--- a/backend/internal/service/billing_service.go
+++ b/backend/internal/service/billing_service.go
@@ -294,6 +294,24 @@ func (s *BillingService) initFallbackPricing() {
 		CacheReadPricePerTokenPriority: 0.35e-6,
 		SupportsCacheBreakdown:         false,
 	}
+	// DeepSeek V4 Pro
+	// Source: https://api-docs.deepseek.com/quick_start/pricing
+	s.fallbackPrices["deepseek-v4-pro"] = &ModelPricing{
+		InputPricePerToken:     4.35e-7,   // $0.435 per MTok
+		OutputPricePerToken:    8.7e-7,    // $0.87 per MTok
+		CacheReadPricePerToken: 3.625e-9,  // $0.003625 per MTok
+		SupportsCacheBreakdown: false,
+	}
+
+	// DeepSeek V4 Flash
+	// Source: https://api-docs.deepseek.com/quick_start/pricing
+	s.fallbackPrices["deepseek-v4-flash"] = &ModelPricing{
+		InputPricePerToken:     1.4e-7,   // $0.14 per MTok
+		OutputPricePerToken:    2.8e-7,   // $0.28 per MTok
+		CacheReadPricePerToken: 2.8e-9,   // $0.0028 per MTok
+		SupportsCacheBreakdown: false,
+	}
+
 	// Codex 族兜底统一按 GPT-5.3 Codex 价格计费
 	s.fallbackPrices["gpt-5.3-codex"] = &ModelPricing{
 		InputPricePerToken:             1.5e-6, // $1.5 per MTok
@@ -344,6 +362,18 @@ func (s *BillingService) getFallbackPricing(model string) *ModelPricing {
 		return s.fallbackPrices["gemini-3.1-pro"]
 	}
 
+	// DeepSeek V4 系列
+	if strings.Contains(modelLower, "deepseek-v4-flash") {
+		return s.fallbackPrices["deepseek-v4-flash"]
+	}
+	if strings.Contains(modelLower, "deepseek-v4-pro") {
+		return s.fallbackPrices["deepseek-v4-pro"]
+	}
+	// deepseek-chat / deepseek-reasoner → V4 Flash（官方兼容别名）
+	if strings.Contains(modelLower, "deepseek-chat") || strings.Contains(modelLower, "deepseek-reasoner") {
+		return s.fallbackPrices["deepseek-v4-flash"]
+	}
+
 	// OpenAI 仅匹配已知 GPT-5/Codex 族，避免未知 OpenAI 型号误计价。
 	if normalized := normalizeKnownOpenAICodexModel(modelLower); normalized != "" {
 		switch normalized {
diff --git a/backend/internal/service/billing_service_test.go b/backend/internal/service/billing_service_test.go
index 52adc31c85..d75bbfde35 100644
--- a/backend/internal/service/billing_service_test.go
+++ b/backend/internal/service/billing_service_test.go
@@ -352,6 +352,10 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 		{name: "openai legacy gpt5.1 codex falls back to gpt5.3 codex", model: "gpt-5.1-codex", expectedInput: 1.5e-6},
 		{name: "openai legacy codex mini latest falls back to gpt5.3 codex", model: "codex-mini-latest", expectedInput: 1.5e-6},
 		{name: "openai unknown no fallback", model: "gpt-unknown-model", expectNilPricing: true},
+		{name: "deepseek v4 pro", model: "deepseek-v4-pro", expectedInput: 4.35e-7},
+		{name: "deepseek v4 flash", model: "deepseek-v4-flash", expectedInput: 1.4e-7},
+		{name: "deepseek chat alias → flash", model: "deepseek-chat", expectedInput: 1.4e-7},
+		{name: "deepseek reasoner alias → flash", model: "deepseek-reasoner", expectedInput: 1.4e-7},
 		{name: "non supported family", model: "qwen-max", expectNilPricing: true},
 	}
 

From 27e26a3a90a52b06ea9f1c64488c153db2183410 Mon Sep 17 00:00:00 2001
From: alfadb <alfadb@163.com>
Date: Sat, 2 May 2026 11:26:41 +0800
Subject: [PATCH 2/8] chore: fix gofmt alignment

---
 backend/internal/service/billing_service.go | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/backend/internal/service/billing_service.go b/backend/internal/service/billing_service.go
index c4e376fcff..bdeef9185b 100644
--- a/backend/internal/service/billing_service.go
+++ b/backend/internal/service/billing_service.go
@@ -297,18 +297,18 @@ func (s *BillingService) initFallbackPricing() {
 	// DeepSeek V4 Pro
 	// Source: https://api-docs.deepseek.com/quick_start/pricing
 	s.fallbackPrices["deepseek-v4-pro"] = &ModelPricing{
-		InputPricePerToken:     4.35e-7,   // $0.435 per MTok
-		OutputPricePerToken:    8.7e-7,    // $0.87 per MTok
-		CacheReadPricePerToken: 3.625e-9,  // $0.003625 per MTok
+		InputPricePerToken:     4.35e-7,  // $0.435 per MTok
+		OutputPricePerToken:    8.7e-7,   // $0.87 per MTok
+		CacheReadPricePerToken: 3.625e-9, // $0.003625 per MTok
 		SupportsCacheBreakdown: false,
 	}
 
 	// DeepSeek V4 Flash
 	// Source: https://api-docs.deepseek.com/quick_start/pricing
 	s.fallbackPrices["deepseek-v4-flash"] = &ModelPricing{
-		InputPricePerToken:     1.4e-7,   // $0.14 per MTok
-		OutputPricePerToken:    2.8e-7,   // $0.28 per MTok
-		CacheReadPricePerToken: 2.8e-9,   // $0.0028 per MTok
+		InputPricePerToken:     1.4e-7, // $0.14 per MTok
+		OutputPricePerToken:    2.8e-7, // $0.28 per MTok
+		CacheReadPricePerToken: 2.8e-9, // $0.0028 per MTok
 		SupportsCacheBreakdown: false,
 	}
 

From 5a593a511e448af9dd4304a9956b81d7f1e2aa7a Mon Sep 17 00:00:00 2001
From: alfadb <alfadb@163.com>
Date: Sat, 9 May 2026 21:13:22 +0800
Subject: [PATCH 3/8] test(billing): tighten DeepSeek V4 fallback assertions;
 clarify branch comments
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Address copilot-pull-request-reviewer feedback on #2157:

- billing_service_test.go: extend TestGetFallbackPricing_FamilyMatching
  with optional expectedOutput / expectedCacheRead fields and assert
  full Input/Output/CacheRead pricing for all 4 DeepSeek cases
  (v4-pro, v4-flash, deepseek-chat→flash, deepseek-reasoner→flash),
  preventing silent regression to 0 output/cache cost.
- billing_service.go: rewrite the DeepSeek block comment to explicitly
  describe its scope (V4 Pro/Flash + chat/reasoner aliases, no
  unknown-deepseek fallback) and tighten the OpenAI comment to make
  it unambiguous that it only describes the OpenAI/Codex branch
  immediately below it.
---
 backend/internal/service/billing_service.go   |  6 +--
 .../internal/service/billing_service_test.go  | 51 ++++++++++++++++---
 2 files changed, 46 insertions(+), 11 deletions(-)

diff --git a/backend/internal/service/billing_service.go b/backend/internal/service/billing_service.go
index bdeef9185b..bbe9e97272 100644
--- a/backend/internal/service/billing_service.go
+++ b/backend/internal/service/billing_service.go
@@ -362,19 +362,19 @@ func (s *BillingService) getFallbackPricing(model string) *ModelPricing {
 		return s.fallbackPrices["gemini-3.1-pro"]
 	}
 
-	// DeepSeek V4 系列
+	// DeepSeek V4 系列：仅匹配已知 V4 Pro/Flash 与官方兼容别名
+	// （deepseek-chat / deepseek-reasoner → V4 Flash），未知 deepseek-* 型号不回退，避免误计价。
 	if strings.Contains(modelLower, "deepseek-v4-flash") {
 		return s.fallbackPrices["deepseek-v4-flash"]
 	}
 	if strings.Contains(modelLower, "deepseek-v4-pro") {
 		return s.fallbackPrices["deepseek-v4-pro"]
 	}
-	// deepseek-chat / deepseek-reasoner → V4 Flash（官方兼容别名）
 	if strings.Contains(modelLower, "deepseek-chat") || strings.Contains(modelLower, "deepseek-reasoner") {
 		return s.fallbackPrices["deepseek-v4-flash"]
 	}
 
-	// OpenAI 仅匹配已知 GPT-5/Codex 族，避免未知 OpenAI 型号误计价。
+	// OpenAI（GPT-5 / Codex 族）：仅匹配已知型号，避免未知 OpenAI 型号误计价。
 	if normalized := normalizeKnownOpenAICodexModel(modelLower); normalized != "" {
 		switch normalized {
 		case "gpt-5.5":
diff --git a/backend/internal/service/billing_service_test.go b/backend/internal/service/billing_service_test.go
index d75bbfde35..3857715b07 100644
--- a/backend/internal/service/billing_service_test.go
+++ b/backend/internal/service/billing_service_test.go
@@ -332,11 +332,14 @@ func TestCalculateCost_LongContextAppliesMultiplierToCacheCreation5mAnd1h(t *tes
 func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 	svc := newTestBillingService()
 
+	// expectedOutput / expectedCacheRead 为 0 时跳过该字段断言（保持与原有用例兼容）。
 	tests := []struct {
-		name             string
-		model            string
-		expectedInput    float64
-		expectNilPricing bool
+		name              string
+		model             string
+		expectedInput     float64
+		expectedOutput    float64
+		expectedCacheRead float64
+		expectNilPricing  bool
 	}{
 		{name: "empty model", model: "   ", expectNilPricing: true},
 		{name: "claude opus 4.6", model: "claude-opus-4.6-20260201", expectedInput: 5e-6},
@@ -352,10 +355,34 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 		{name: "openai legacy gpt5.1 codex falls back to gpt5.3 codex", model: "gpt-5.1-codex", expectedInput: 1.5e-6},
 		{name: "openai legacy codex mini latest falls back to gpt5.3 codex", model: "codex-mini-latest", expectedInput: 1.5e-6},
 		{name: "openai unknown no fallback", model: "gpt-unknown-model", expectNilPricing: true},
-		{name: "deepseek v4 pro", model: "deepseek-v4-pro", expectedInput: 4.35e-7},
-		{name: "deepseek v4 flash", model: "deepseek-v4-flash", expectedInput: 1.4e-7},
-		{name: "deepseek chat alias → flash", model: "deepseek-chat", expectedInput: 1.4e-7},
-		{name: "deepseek reasoner alias → flash", model: "deepseek-reasoner", expectedInput: 1.4e-7},
+		{
+			name:              "deepseek v4 pro",
+			model:             "deepseek-v4-pro",
+			expectedInput:     4.35e-7,
+			expectedOutput:    8.7e-7,
+			expectedCacheRead: 3.625e-9,
+		},
+		{
+			name:              "deepseek v4 flash",
+			model:             "deepseek-v4-flash",
+			expectedInput:     1.4e-7,
+			expectedOutput:    2.8e-7,
+			expectedCacheRead: 2.8e-9,
+		},
+		{
+			name:              "deepseek chat alias → flash",
+			model:             "deepseek-chat",
+			expectedInput:     1.4e-7,
+			expectedOutput:    2.8e-7,
+			expectedCacheRead: 2.8e-9,
+		},
+		{
+			name:              "deepseek reasoner alias → flash",
+			model:             "deepseek-reasoner",
+			expectedInput:     1.4e-7,
+			expectedOutput:    2.8e-7,
+			expectedCacheRead: 2.8e-9,
+		},
 		{name: "non supported family", model: "qwen-max", expectNilPricing: true},
 	}
 
@@ -368,6 +395,14 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			}
 			require.NotNil(t, pricing)
 			require.InDelta(t, tt.expectedInput, pricing.InputPricePerToken, 1e-12)
+			if tt.expectedOutput != 0 {
+				require.InDelta(t, tt.expectedOutput, pricing.OutputPricePerToken, 1e-12,
+					"OutputPricePerToken mismatch for %s", tt.model)
+			}
+			if tt.expectedCacheRead != 0 {
+				require.InDelta(t, tt.expectedCacheRead, pricing.CacheReadPricePerToken, 1e-14,
+					"CacheReadPricePerToken mismatch for %s", tt.model)
+			}
 		})
 	}
 }

From f597d98bdf343bcd4c682c20fcd070cc9af51560 Mon Sep 17 00:00:00 2001
From: alfadb <alfadb@163.com>
Date: Fri, 15 May 2026 09:31:04 +0800
Subject: [PATCH 4/8] test(openai): use unpriced model in usage test

---
 .../internal/service/openai_gateway_record_usage_test.go    | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/backend/internal/service/openai_gateway_record_usage_test.go b/backend/internal/service/openai_gateway_record_usage_test.go
index 1683fa4279..1e4e58dc42 100644
--- a/backend/internal/service/openai_gateway_record_usage_test.go
+++ b/backend/internal/service/openai_gateway_record_usage_test.go
@@ -327,7 +327,7 @@ func TestOpenAIGatewayServiceRecordUsage_MissingPricingRecordsZeroCostUsageLog(t
 				InputTokens:  1200,
 				OutputTokens: 300,
 			},
-			Model:    "deepseek-v4-flash",
+			Model:    "pricing-missing-test-model",
 			Duration: time.Second,
 		},
 		APIKey:        &APIKey{ID: 1002, Quota: 100, Group: &Group{RateMultiplier: 1}},
@@ -346,8 +346,8 @@ func TestOpenAIGatewayServiceRecordUsage_MissingPricingRecordsZeroCostUsageLog(t
 
 	require.NotNil(t, usageRepo.lastLog)
 	require.Equal(t, "resp_missing_pricing", usageRepo.lastLog.RequestID)
-	require.Equal(t, "deepseek-v4-flash", usageRepo.lastLog.Model)
-	require.Equal(t, "deepseek-v4-flash", usageRepo.lastLog.RequestedModel)
+	require.Equal(t, "pricing-missing-test-model", usageRepo.lastLog.Model)
+	require.Equal(t, "pricing-missing-test-model", usageRepo.lastLog.RequestedModel)
 	require.Equal(t, 1200, usageRepo.lastLog.InputTokens)
 	require.Equal(t, 300, usageRepo.lastLog.OutputTokens)
 	require.Zero(t, usageRepo.lastLog.TotalCost)

From a4ce73391b1491f0e88291425dc5cbbe2af910c1 Mon Sep 17 00:00:00 2001
From: alfadb <alfadb@163.com>
Date: Fri, 12 Jun 2026 10:07:35 +0800
Subject: [PATCH 5/8] feat(billing): add GLM / Kimi / MiniMax fallback pricing
 for Chinese LLM providers
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 fix/deepseek-fallback-pricing 的 DeepSeek V4 Pro/Flash 定价收编，
并扩展国产 LLM 兜底定价覆盖：

- 智谱 GLM (z.ai 公开 SKU 13 个): glm-5.1 / glm-5 / glm-5-turbo / glm-4.7 /
  glm-4.7-flashx / glm-4.6 / glm-4.5 / glm-4.5-x / glm-4.5-air / glm-4.5-airx /
  glm-4-32b-0414-128k / glm-4.5-flash / glm-4.7-flash
- 月之暗面 Kimi K 系列 4 个: kimi-k2.6 / kimi-k2.5 / kimi-k2-thinking /
  kimi-k2 (K2-0905/K2-0711 官方未保留定价，隐性回退到 kimi-k2)
- MiniMax M 系列 6 个: minimax-m3 (≤512K) / minimax-m2.7 / minimax-m2.7-highspeed /
  minimax-m2.5 / minimax-m2.1 / minimax-m2
- DeepSeek V4 沿用原 fix/deepseek-fallback-pricing 的 entry 移植过来

所有定价数据来自各家官方定价页 (USD/MTok 口径)，与现有
Claude/GPT 风格保持一致 (硬编码在 Go 代码里，admin channel 优先)。

匹配策略：长 key 优先 (glm-5.1 优先于 glm-5, k2.6 优先于 k2)；
未列出的国产厂商 alias (qwen/doubao/hunyuan) 一律不返回兜底价，
避免误计价——这与 fix/deepseek-fallback-pricing 的白名单语义一致。

数据源注释 (Source URL) 内嵌在每个 initFallbackPricing 分组开头，
便于后续 PR 同步上游价格变化。
---
 backend/internal/service/billing_service.go   | 255 ++++++++++++++++++
 .../internal/service/billing_service_test.go  | 206 +++++++++++++-
 2 files changed, 460 insertions(+), 1 deletion(-)

diff --git a/backend/internal/service/billing_service.go b/backend/internal/service/billing_service.go
index bbe9e97272..a1008e0515 100644
--- a/backend/internal/service/billing_service.go
+++ b/backend/internal/service/billing_service.go
@@ -323,6 +323,180 @@ func (s *BillingService) initFallbackPricing() {
 		CacheReadPricePerTokenPriority: 0.3e-6,
 		SupportsCacheBreakdown:         false,
 	}
+
+	// ============================================================
+	// 国产 LLM 兜底定价（数据源：各家官方定价页/USD 口径）
+	// 顺序：DeepSeek → 智谱 GLM → 月之暗面 Kimi → MiniMax
+	// 覆盖逻辑见同文件 getFallbackPricing()
+	// ============================================================
+
+	// ---- DeepSeek V4 系列 ----
+	// Source: https://api-docs.deepseek.com/quick_start/pricing
+	// （deepseek-chat / deepseek-reasoner 为 deepseek-v4-flash 的兼容别名，2026/07/24 弃用）
+	s.fallbackPrices["deepseek-v4-pro"] = &ModelPricing{
+		InputPricePerToken:     4.35e-7,  // $0.435 per MTok (cache miss)
+		OutputPricePerToken:    8.7e-7,   // $0.87 per MTok
+		CacheReadPricePerToken: 3.625e-9, // $0.003625 per MTok (cache hit)
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["deepseek-v4-flash"] = &ModelPricing{
+		InputPricePerToken:     1.4e-7, // $0.14 per MTok (cache miss)
+		OutputPricePerToken:    2.8e-7, // $0.28 per MTok
+		CacheReadPricePerToken: 2.8e-9, // $0.0028 per MTok (cache hit)
+		SupportsCacheBreakdown: false,
+	}
+
+	// ---- 智谱 GLM（Z.AI）----
+	// Source: https://docs.z.ai/guides/overview/pricing (USD per 1M tokens)
+	// 注意：CacheReadPricePerToken 即"缓存命中"价格，CacheCreationPricePerToken 留空（智谱未公开写入价，按 0 处理）。
+	// GLM-4.6 与 GLM-4.5 在 z.ai 国际版上定价一致；GLM-4.5 国内按 ¥0.8/¥2，汇率换算后约 $0.112/$0.28，与国际版 $0.6/$2.2 不同，本分支采用国际版 USD 口径与现有 Claude/GPT 一致。
+	s.fallbackPrices["glm-5.1"] = &ModelPricing{
+		InputPricePerToken:     1.4e-6, // $1.40 per MTok
+		OutputPricePerToken:    4.4e-6, // $4.40 per MTok
+		CacheReadPricePerToken: 0.26e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-5"] = &ModelPricing{
+		InputPricePerToken:     1e-6, // $1.00 per MTok
+		OutputPricePerToken:    3.2e-6,
+		CacheReadPricePerToken: 0.2e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-5-turbo"] = &ModelPricing{
+		InputPricePerToken:     1.2e-6,
+		OutputPricePerToken:    4e-6,
+		CacheReadPricePerToken: 0.24e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4.7"] = &ModelPricing{
+		InputPricePerToken:     0.6e-6, // $0.60 per MTok
+		OutputPricePerToken:    2.2e-6,
+		CacheReadPricePerToken: 0.11e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4.7-flashx"] = &ModelPricing{
+		InputPricePerToken:     0.07e-6, // $0.07 per MTok
+		OutputPricePerToken:    0.4e-6,
+		CacheReadPricePerToken: 0.01e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4.6"] = &ModelPricing{
+		InputPricePerToken:     0.6e-6, // $0.60 per MTok
+		OutputPricePerToken:    2.2e-6,
+		CacheReadPricePerToken: 0.11e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4.5"] = &ModelPricing{
+		InputPricePerToken:     0.6e-6, // $0.60 per MTok
+		OutputPricePerToken:    2.2e-6,
+		CacheReadPricePerToken: 0.11e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4.5-x"] = &ModelPricing{
+		InputPricePerToken:     2.2e-6, // $2.20 per MTok
+		OutputPricePerToken:    8.9e-6,
+		CacheReadPricePerToken: 0.45e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4.5-air"] = &ModelPricing{
+		InputPricePerToken:     0.2e-6, // $0.20 per MTok
+		OutputPricePerToken:    1.1e-6,
+		CacheReadPricePerToken: 0.03e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4.5-airx"] = &ModelPricing{
+		InputPricePerToken:     1.1e-6,
+		OutputPricePerToken:    4.5e-6,
+		CacheReadPricePerToken: 0.22e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4-32b-0414-128k"] = &ModelPricing{
+		InputPricePerToken:     0.1e-6, // $0.10 per MTok
+		OutputPricePerToken:    0.1e-6,
+		SupportsCacheBreakdown: false,
+	}
+	// GLM-4.5-Flash / GLM-4.7-Flash 在 z.ai 上为 Free，保留 zero-cost entry 防止未知 alias 误计费。
+	s.fallbackPrices["glm-4.5-flash"] = &ModelPricing{
+		InputPricePerToken:     0,
+		OutputPricePerToken:    0,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["glm-4.7-flash"] = &ModelPricing{
+		InputPricePerToken:     0,
+		OutputPricePerToken:    0,
+		SupportsCacheBreakdown: false,
+	}
+
+	// ---- 月之暗面 Kimi（K 系列）----
+	// Source: https://platform.moonshot.cn/docs/pricing/overview (元/百万 tokens 口径)
+	//       交叉验证：https://www.tmtpost.com/7961404.html (USD 口径)
+	// Moonshot V1 (¥2/¥5/¥10 多 tier) 公开页未直接标注 USD 价，本分支不覆盖，避免误计价。
+	// K2-0905 / K2-0711 官方页面未保留定价，不覆盖。
+	s.fallbackPrices["kimi-k2.6"] = &ModelPricing{
+		InputPricePerToken:     0.95e-6, // $0.95 per MTok (cache miss)
+		OutputPricePerToken:    4e-6,    // $4.00 per MTok
+		CacheReadPricePerToken: 0.15e-6, // $0.15 per MTok (cache hit, ¥1.10)
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["kimi-k2.5"] = &ModelPricing{
+		InputPricePerToken:     0.60e-6, // $0.60 per MTok
+		OutputPricePerToken:    3e-6,    // $3.00 per MTok
+		CacheReadPricePerToken: 0.098e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["kimi-k2-thinking"] = &ModelPricing{
+		InputPricePerToken:     0.56e-6, // ¥4/百万 ≈ $0.56
+		OutputPricePerToken:    2.24e-6, // ¥16/百万
+		CacheReadPricePerToken: 0.14e-6, // ¥1/百万
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["kimi-k2"] = &ModelPricing{
+		InputPricePerToken:     0.56e-6, // ¥4/百万
+		OutputPricePerToken:    2.24e-6, // ¥16/百万
+		CacheReadPricePerToken: 0.14e-6, // ¥1/百万
+		SupportsCacheBreakdown: false,
+	}
+
+	// ---- MiniMax M 系列 ----
+	// Source: https://platform.minimax.io/docs/guides/pricing-paygo
+	// 注意：MiniMax M3 在 >512K context 时价格翻倍，本兜底采用 ≤512K 标准 tier（保守口径，对用户有利）。
+	// 如需支持长上下文 multiplier，可后续参考 GPT-5.4 模式扩展 LongContextXxx 字段。
+	s.fallbackPrices["minimax-m3"] = &ModelPricing{
+		InputPricePerToken:     0.60e-6, // $0.60 per MTok (≤512K standard tier, 含 50% 永久折扣前原价 $1.20)
+		OutputPricePerToken:    2.40e-6,
+		CacheReadPricePerToken: 0.12e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["minimax-m2.7"] = &ModelPricing{
+		InputPricePerToken:     0.30e-6, // $0.30 per MTok
+		OutputPricePerToken:    1.20e-6,
+		CacheReadPricePerToken: 0.06e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["minimax-m2.7-highspeed"] = &ModelPricing{
+		InputPricePerToken:     0.60e-6,
+		OutputPricePerToken:    2.40e-6,
+		CacheReadPricePerToken: 0.06e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["minimax-m2.5"] = &ModelPricing{
+		InputPricePerToken:     0.30e-6,
+		OutputPricePerToken:    1.20e-6,
+		CacheReadPricePerToken: 0.03e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["minimax-m2.1"] = &ModelPricing{
+		InputPricePerToken:     0.30e-6,
+		OutputPricePerToken:    1.20e-6,
+		CacheReadPricePerToken: 0.03e-6,
+		SupportsCacheBreakdown: false,
+	}
+	s.fallbackPrices["minimax-m2"] = &ModelPricing{
+		InputPricePerToken:     0.30e-6,
+		OutputPricePerToken:    1.20e-6,
+		CacheReadPricePerToken: 0.03e-6,
+		SupportsCacheBreakdown: false,
+	}
 }
 
 // getFallbackPricing 根据模型系列获取回退价格
@@ -374,6 +548,87 @@ func (s *BillingService) getFallbackPricing(model string) *ModelPricing {
 		return s.fallbackPrices["deepseek-v4-flash"]
 	}
 
+	// ---- 国产 LLM 兜底匹配 ----
+	// 匹配策略：长 key 优先（具体模型 → 系列 / 厂商），未知型号不回退以避免误计价。
+	// 与 DeepSeek 一样采用"白名单"语义：未在本表命中的国产模型 alias 一律不返回兜底价。
+
+	// 智谱 GLM（z.ai 公开 SKU：glm-5.1 / glm-5 / glm-5-turbo / glm-4.7 / glm-4.6 / glm-4.5 等）
+	// 匹配顺序：先判别最高 tier，再依次降级。
+	if strings.Contains(modelLower, "glm-5.1") {
+		return s.fallbackPrices["glm-5.1"]
+	}
+	if strings.Contains(modelLower, "glm-5-turbo") || strings.Contains(modelLower, "glm-5turbo") {
+		return s.fallbackPrices["glm-5-turbo"]
+	}
+	if strings.Contains(modelLower, "glm-5") {
+		return s.fallbackPrices["glm-5"]
+	}
+	if strings.Contains(modelLower, "glm-4.7-flashx") {
+		return s.fallbackPrices["glm-4.7-flashx"]
+	}
+	if strings.Contains(modelLower, "glm-4.7-flash") {
+		return s.fallbackPrices["glm-4.7-flash"]
+	}
+	if strings.Contains(modelLower, "glm-4.7") {
+		return s.fallbackPrices["glm-4.7"]
+	}
+	if strings.Contains(modelLower, "glm-4.6") {
+		return s.fallbackPrices["glm-4.6"]
+	}
+	if strings.Contains(modelLower, "glm-4.5-flash") {
+		return s.fallbackPrices["glm-4.5-flash"]
+	}
+	if strings.Contains(modelLower, "glm-4.5-x") || strings.Contains(modelLower, "glm-4.5x") {
+		return s.fallbackPrices["glm-4.5-x"]
+	}
+	if strings.Contains(modelLower, "glm-4.5-airx") || strings.Contains(modelLower, "glm-4.5airx") {
+		return s.fallbackPrices["glm-4.5-airx"]
+	}
+	if strings.Contains(modelLower, "glm-4.5-air") || strings.Contains(modelLower, "glm-4.5air") {
+		return s.fallbackPrices["glm-4.5-air"]
+	}
+	if strings.Contains(modelLower, "glm-4.5") {
+		return s.fallbackPrices["glm-4.5"]
+	}
+	if strings.Contains(modelLower, "glm-4-32b") {
+		return s.fallbackPrices["glm-4-32b-0414-128k"]
+	}
+
+	// 月之暗面 Kimi（kimi-k2.6 / kimi-k2.5 / kimi-k2-thinking / kimi-k2）
+	// K2-0905 / K2-0711 官方未保留定价，不进入 fallback。
+	if strings.Contains(modelLower, "kimi-k2.6") || strings.Contains(modelLower, "kimi-k2-6") {
+		return s.fallbackPrices["kimi-k2.6"]
+	}
+	if strings.Contains(modelLower, "kimi-k2.5") || strings.Contains(modelLower, "kimi-k2-5") {
+		return s.fallbackPrices["kimi-k2.5"]
+	}
+	if strings.Contains(modelLower, "kimi-k2-thinking") || strings.Contains(modelLower, "kimi-k2-thinking-") {
+		return s.fallbackPrices["kimi-k2-thinking"]
+	}
+	if strings.Contains(modelLower, "kimi-k2") || strings.Contains(modelLower, "kimi/k2") {
+		return s.fallbackPrices["kimi-k2"]
+	}
+
+	// MiniMax M 系列（M3 / M2.7 / M2.5 / M2.1 / M2；含 highspeed 变体）
+	if strings.Contains(modelLower, "minimax-m3") || strings.Contains(modelLower, "MiniMax-M3") {
+		return s.fallbackPrices["minimax-m3"]
+	}
+	if strings.Contains(modelLower, "minimax-m2.7-highspeed") || strings.Contains(modelLower, "minimax-m2-7-highspeed") {
+		return s.fallbackPrices["minimax-m2.7-highspeed"]
+	}
+	if strings.Contains(modelLower, "minimax-m2.7") || strings.Contains(modelLower, "minimax-m2-7") {
+		return s.fallbackPrices["minimax-m2.7"]
+	}
+	if strings.Contains(modelLower, "minimax-m2.5") || strings.Contains(modelLower, "minimax-m2-5") {
+		return s.fallbackPrices["minimax-m2.5"]
+	}
+	if strings.Contains(modelLower, "minimax-m2.1") || strings.Contains(modelLower, "minimax-m2-1") {
+		return s.fallbackPrices["minimax-m2.1"]
+	}
+	if strings.Contains(modelLower, "minimax-m2") || strings.Contains(modelLower, "minimax-m-2") {
+		return s.fallbackPrices["minimax-m2"]
+	}
+
 	// OpenAI（GPT-5 / Codex 族）：仅匹配已知型号，避免未知 OpenAI 型号误计价。
 	if normalized := normalizeKnownOpenAICodexModel(modelLower); normalized != "" {
 		switch normalized {
diff --git a/backend/internal/service/billing_service_test.go b/backend/internal/service/billing_service_test.go
index 3857715b07..4935ad8487 100644
--- a/backend/internal/service/billing_service_test.go
+++ b/backend/internal/service/billing_service_test.go
@@ -383,7 +383,211 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			expectedOutput:    2.8e-7,
 			expectedCacheRead: 2.8e-9,
 		},
-		{name: "non supported family", model: "qwen-max", expectNilPricing: true},
+
+		// ---- 智谱 GLM（z.ai USD 口径）----
+		{
+			name:              "glm 5.1 flagship",
+			model:             "glm-5.1",
+			expectedInput:     1.4e-6,
+			expectedOutput:    4.4e-6,
+			expectedCacheRead: 0.26e-6,
+		},
+		{
+			name:              "glm 5 base",
+			model:             "glm-5",
+			expectedInput:     1e-6,
+			expectedOutput:    3.2e-6,
+			expectedCacheRead: 0.2e-6,
+		},
+		{
+			name:              "glm 5 turbo",
+			model:             "glm-5-turbo",
+			expectedInput:     1.2e-6,
+			expectedOutput:    4e-6,
+			expectedCacheRead: 0.24e-6,
+		},
+		{
+			name:              "glm 4.7",
+			model:             "glm-4.7",
+			expectedInput:     0.6e-6,
+			expectedOutput:    2.2e-6,
+			expectedCacheRead: 0.11e-6,
+		},
+		{
+			name:              "glm 4.6",
+			model:             "glm-4.6",
+			expectedInput:     0.6e-6,
+			expectedOutput:    2.2e-6,
+			expectedCacheRead: 0.11e-6,
+		},
+		{
+			name:              "glm 4.5",
+			model:             "glm-4.5",
+			expectedInput:     0.6e-6,
+			expectedOutput:    2.2e-6,
+			expectedCacheRead: 0.11e-6,
+		},
+		{
+			name:              "glm 4.5-x premium",
+			model:             "glm-4.5-x",
+			expectedInput:     2.2e-6,
+			expectedOutput:    8.9e-6,
+			expectedCacheRead: 0.45e-6,
+		},
+		{
+			name:              "glm 4.5-air lightweight",
+			model:             "glm-4.5-air",
+			expectedInput:     0.2e-6,
+			expectedOutput:    1.1e-6,
+			expectedCacheRead: 0.03e-6,
+		},
+		{
+			name:              "glm 4.7-flashx",
+			model:             "glm-4.7-flashx",
+			expectedInput:     0.07e-6,
+			expectedOutput:    0.4e-6,
+			expectedCacheRead: 0.01e-6,
+		},
+		{
+			name:              "glm 4.5-flash free tier",
+			model:             "glm-4.5-flash",
+			expectedInput:     0, // Free tier on z.ai
+			expectedOutput:    0,
+			expectedCacheRead: 0,
+		},
+		{
+			name:              "glm 4.7-flash free tier",
+			model:             "glm-4.7-flash",
+			expectedInput:     0,
+			expectedOutput:    0,
+			expectedCacheRead: 0,
+		},
+		{
+			name:              "glm 4-32b legacy",
+			model:             "glm-4-32b-0414-128k",
+			expectedInput:     0.1e-6,
+			expectedOutput:    0.1e-6,
+			expectedCacheRead: 0,
+		},
+		// 关键：5.1 必须先于 5 匹配（避免被 glm-5 抢走）
+		{
+			name:              "glm 5.1 vs glm 5 ordering (verbatim 5.1)",
+			model:             "glm-5.1",
+			expectedInput:     1.4e-6, // = glm-5.1 价格
+			expectedOutput:    4.4e-6,
+			expectedCacheRead: 0.26e-6,
+		},
+		{
+			name:              "glm 4.5-air vs glm 4.5 ordering",
+			model:             "glm-4.5-air",
+			expectedInput:     0.2e-6, // = glm-4.5-air 价格（不是 glm-4.5 的 0.6e-6）
+			expectedOutput:    1.1e-6,
+			expectedCacheRead: 0.03e-6,
+		},
+
+		// ---- 月之暗面 Kimi ----
+		{
+			name:              "kimi k2.6 flagship",
+			model:             "kimi-k2.6",
+			expectedInput:     0.95e-6,
+			expectedOutput:    4e-6,
+			expectedCacheRead: 0.15e-6,
+		},
+		{
+			name:              "kimi k2.5",
+			model:             "kimi-k2.5",
+			expectedInput:     0.60e-6,
+			expectedOutput:    3e-6,
+			expectedCacheRead: 0.098e-6,
+		},
+		{
+			name:              "kimi k2-thinking",
+			model:             "kimi-k2-thinking",
+			expectedInput:     0.56e-6,
+			expectedOutput:    2.24e-6,
+			expectedCacheRead: 0.14e-6,
+		},
+		{
+			name:              "kimi k2 base",
+			model:             "kimi-k2",
+			expectedInput:     0.56e-6,
+			expectedOutput:    2.24e-6,
+			expectedCacheRead: 0.14e-6,
+		},
+		// 关键：k2.6 / k2.5 / k2-thinking 必须先于 k2 匹配
+		{
+			name:              "kimi k2.6 vs k2 ordering",
+			model:             "kimi-k2.6",
+			expectedInput:     0.95e-6, // = k2.6 不是 k2 的 0.56e-6
+			expectedOutput:    4e-6,
+			expectedCacheRead: 0.15e-6,
+		},
+		{
+			name:              "kimi k2 thinking hyphenated variant",
+			model:             "kimi-k2-thinking-preview",
+			expectedInput:     0.56e-6,
+			expectedOutput:    2.24e-6,
+			expectedCacheRead: 0.14e-6,
+		},
+
+		// ---- MiniMax M 系列 ----
+		{
+			name:              "minimax m3",
+			model:             "minimax-m3",
+			expectedInput:     0.60e-6,
+			expectedOutput:    2.40e-6,
+			expectedCacheRead: 0.12e-6,
+		},
+		{
+			name:              "minimax m3 long ctx boundary keep standard tier",
+			model:             "minimax-m3-long", // 仍按 standard tier (≤512K)
+			expectedInput:     0.60e-6,
+			expectedOutput:    2.40e-6,
+			expectedCacheRead: 0.12e-6,
+		},
+		{
+			name:              "minimax m2.7",
+			model:             "minimax-m2.7",
+			expectedInput:     0.30e-6,
+			expectedOutput:    1.20e-6,
+			expectedCacheRead: 0.06e-6,
+		},
+		{
+			name:              "minimax m2.7 highspeed",
+			model:             "minimax-m2.7-highspeed",
+			expectedInput:     0.60e-6,
+			expectedOutput:    2.40e-6,
+			expectedCacheRead: 0.06e-6,
+		},
+		{
+			name:              "minimax m2.5",
+			model:             "minimax-m2.5",
+			expectedInput:     0.30e-6,
+			expectedOutput:    1.20e-6,
+			expectedCacheRead: 0.03e-6,
+		},
+		{
+			name:              "minimax m2 legacy",
+			model:             "minimax-m2",
+			expectedInput:     0.30e-6,
+			expectedOutput:    1.20e-6,
+			expectedCacheRead: 0.03e-6,
+		},
+
+		// ---- 负向用例：未覆盖的国产厂商 / alias 不应误计价 ----
+		{name: "qwen unknown no fallback", model: "qwen-max", expectNilPricing: true},
+		{name: "doubao unknown no fallback", model: "doubao-pro", expectNilPricing: true},
+		{name: "hunyuan unknown no fallback", model: "hunyuan-t1", expectNilPricing: true},
+		{name: "moonshot v1 not covered", model: "moonshot-v1-8k", expectNilPricing: true},
+		// kimi-k2-0905 / kimi-k2-0711 官方未公布独立价，走 kimi-k2 隐性回退（接受）——
+		// 如未来官方公布独立价，需在 getFallbackPricing 加显式分支。
+		{
+			name:              "kimi k2-0905-preview implicit fallback to k2",
+			model:             "kimi-k2-0905-preview",
+			expectedInput:     0.56e-6,
+			expectedOutput:    2.24e-6,
+			expectedCacheRead: 0.14e-6,
+		},
 	}
 
 	for _, tt := range tests {

From c90089c8143d9cee3c0291316d89dde0590ea75c Mon Sep 17 00:00:00 2001
From: alfadb <alfadb@163.com>
Date: Fri, 12 Jun 2026 10:29:24 +0800
Subject: [PATCH 6/8] fix(billing): address Copilot review feedback
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

应 PR #3243 review 修正三项问题：

1. 删除 DeepSeek V4 fallback 定价的重复块（cherry-pick 时
   的旧位置 + 新国产 LLM 分段都写了一遍，后写覆盖前写，
   当前值相同所以无 bug，但冗余，未来改一处易漏另一处）。
2. 删除 minimax-m3 匹配中的死代码——modelLower 已经
   strings.ToLower()，'MiniMax-M3' 字面量永远不可能匹中。
3. 测试结构改 *float64 替代 0 sentinel，让 free-tier
   模型（GLM-4.5-Flash / GLM-4.7-Flash）的 0 价能被
   真正断言而非默默跳过；定义内联 floatPtr 辅助函数。
---
 backend/internal/service/billing_service.go   |  20 +--
 .../internal/service/billing_service_test.go  | 147 +++++++++---------
 2 files changed, 75 insertions(+), 92 deletions(-)

diff --git a/backend/internal/service/billing_service.go b/backend/internal/service/billing_service.go
index a1008e0515..43e91ce3e2 100644
--- a/backend/internal/service/billing_service.go
+++ b/backend/internal/service/billing_service.go
@@ -294,24 +294,6 @@ func (s *BillingService) initFallbackPricing() {
 		CacheReadPricePerTokenPriority: 0.35e-6,
 		SupportsCacheBreakdown:         false,
 	}
-	// DeepSeek V4 Pro
-	// Source: https://api-docs.deepseek.com/quick_start/pricing
-	s.fallbackPrices["deepseek-v4-pro"] = &ModelPricing{
-		InputPricePerToken:     4.35e-7,  // $0.435 per MTok
-		OutputPricePerToken:    8.7e-7,   // $0.87 per MTok
-		CacheReadPricePerToken: 3.625e-9, // $0.003625 per MTok
-		SupportsCacheBreakdown: false,
-	}
-
-	// DeepSeek V4 Flash
-	// Source: https://api-docs.deepseek.com/quick_start/pricing
-	s.fallbackPrices["deepseek-v4-flash"] = &ModelPricing{
-		InputPricePerToken:     1.4e-7, // $0.14 per MTok
-		OutputPricePerToken:    2.8e-7, // $0.28 per MTok
-		CacheReadPricePerToken: 2.8e-9, // $0.0028 per MTok
-		SupportsCacheBreakdown: false,
-	}
-
 	// Codex 族兜底统一按 GPT-5.3 Codex 价格计费
 	s.fallbackPrices["gpt-5.3-codex"] = &ModelPricing{
 		InputPricePerToken:             1.5e-6, // $1.5 per MTok
@@ -610,7 +592,7 @@ func (s *BillingService) getFallbackPricing(model string) *ModelPricing {
 	}
 
 	// MiniMax M 系列（M3 / M2.7 / M2.5 / M2.1 / M2；含 highspeed 变体）
-	if strings.Contains(modelLower, "minimax-m3") || strings.Contains(modelLower, "MiniMax-M3") {
+	if strings.Contains(modelLower, "minimax-m3") {
 		return s.fallbackPrices["minimax-m3"]
 	}
 	if strings.Contains(modelLower, "minimax-m2.7-highspeed") || strings.Contains(modelLower, "minimax-m2-7-highspeed") {
diff --git a/backend/internal/service/billing_service_test.go b/backend/internal/service/billing_service_test.go
index 4935ad8487..8bd6e41ede 100644
--- a/backend/internal/service/billing_service_test.go
+++ b/backend/internal/service/billing_service_test.go
@@ -332,13 +332,15 @@ func TestCalculateCost_LongContextAppliesMultiplierToCacheCreation5mAnd1h(t *tes
 func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 	svc := newTestBillingService()
 
-	// expectedOutput / expectedCacheRead 为 0 时跳过该字段断言（保持与原有用例兼容）。
+	floatPtr := func(v float64) *float64 { return &v }
+
+	// expectedOutput / expectedCacheRead 为 nil 时跳过该字段断言（保持与原有用例兼容）。
 	tests := []struct {
 		name              string
 		model             string
 		expectedInput     float64
-		expectedOutput    float64
-		expectedCacheRead float64
+		expectedOutput    *float64
+		expectedCacheRead *float64
 		expectNilPricing  bool
 	}{
 		{name: "empty model", model: "   ", expectNilPricing: true},
@@ -359,29 +361,29 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			name:              "deepseek v4 pro",
 			model:             "deepseek-v4-pro",
 			expectedInput:     4.35e-7,
-			expectedOutput:    8.7e-7,
-			expectedCacheRead: 3.625e-9,
+			expectedOutput:    floatPtr(8.7e-7),
+			expectedCacheRead: floatPtr(3.625e-9),
 		},
 		{
 			name:              "deepseek v4 flash",
 			model:             "deepseek-v4-flash",
 			expectedInput:     1.4e-7,
-			expectedOutput:    2.8e-7,
-			expectedCacheRead: 2.8e-9,
+			expectedOutput:    floatPtr(2.8e-7),
+			expectedCacheRead: floatPtr(2.8e-9),
 		},
 		{
 			name:              "deepseek chat alias → flash",
 			model:             "deepseek-chat",
 			expectedInput:     1.4e-7,
-			expectedOutput:    2.8e-7,
-			expectedCacheRead: 2.8e-9,
+			expectedOutput:    floatPtr(2.8e-7),
+			expectedCacheRead: floatPtr(2.8e-9),
 		},
 		{
 			name:              "deepseek reasoner alias → flash",
 			model:             "deepseek-reasoner",
 			expectedInput:     1.4e-7,
-			expectedOutput:    2.8e-7,
-			expectedCacheRead: 2.8e-9,
+			expectedOutput:    floatPtr(2.8e-7),
+			expectedCacheRead: floatPtr(2.8e-9),
 		},
 
 		// ---- 智谱 GLM（z.ai USD 口径）----
@@ -389,100 +391,99 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			name:              "glm 5.1 flagship",
 			model:             "glm-5.1",
 			expectedInput:     1.4e-6,
-			expectedOutput:    4.4e-6,
-			expectedCacheRead: 0.26e-6,
+			expectedOutput:    floatPtr(4.4e-6),
+			expectedCacheRead: floatPtr(0.26e-6),
 		},
 		{
 			name:              "glm 5 base",
 			model:             "glm-5",
 			expectedInput:     1e-6,
-			expectedOutput:    3.2e-6,
-			expectedCacheRead: 0.2e-6,
+			expectedOutput:    floatPtr(3.2e-6),
+			expectedCacheRead: floatPtr(0.2e-6),
 		},
 		{
 			name:              "glm 5 turbo",
 			model:             "glm-5-turbo",
 			expectedInput:     1.2e-6,
-			expectedOutput:    4e-6,
-			expectedCacheRead: 0.24e-6,
+			expectedOutput:    floatPtr(4e-6),
+			expectedCacheRead: floatPtr(0.24e-6),
 		},
 		{
 			name:              "glm 4.7",
 			model:             "glm-4.7",
 			expectedInput:     0.6e-6,
-			expectedOutput:    2.2e-6,
-			expectedCacheRead: 0.11e-6,
+			expectedOutput:    floatPtr(2.2e-6),
+			expectedCacheRead: floatPtr(0.11e-6),
 		},
 		{
 			name:              "glm 4.6",
 			model:             "glm-4.6",
 			expectedInput:     0.6e-6,
-			expectedOutput:    2.2e-6,
-			expectedCacheRead: 0.11e-6,
+			expectedOutput:    floatPtr(2.2e-6),
+			expectedCacheRead: floatPtr(0.11e-6),
 		},
 		{
 			name:              "glm 4.5",
 			model:             "glm-4.5",
 			expectedInput:     0.6e-6,
-			expectedOutput:    2.2e-6,
-			expectedCacheRead: 0.11e-6,
+			expectedOutput:    floatPtr(2.2e-6),
+			expectedCacheRead: floatPtr(0.11e-6),
 		},
 		{
 			name:              "glm 4.5-x premium",
 			model:             "glm-4.5-x",
 			expectedInput:     2.2e-6,
-			expectedOutput:    8.9e-6,
-			expectedCacheRead: 0.45e-6,
+			expectedOutput:    floatPtr(8.9e-6),
+			expectedCacheRead: floatPtr(0.45e-6),
 		},
 		{
 			name:              "glm 4.5-air lightweight",
 			model:             "glm-4.5-air",
 			expectedInput:     0.2e-6,
-			expectedOutput:    1.1e-6,
-			expectedCacheRead: 0.03e-6,
+			expectedOutput:    floatPtr(1.1e-6),
+			expectedCacheRead: floatPtr(0.03e-6),
 		},
 		{
 			name:              "glm 4.7-flashx",
 			model:             "glm-4.7-flashx",
 			expectedInput:     0.07e-6,
-			expectedOutput:    0.4e-6,
-			expectedCacheRead: 0.01e-6,
+			expectedOutput:    floatPtr(0.4e-6),
+			expectedCacheRead: floatPtr(0.01e-6),
 		},
 		{
 			name:              "glm 4.5-flash free tier",
 			model:             "glm-4.5-flash",
 			expectedInput:     0, // Free tier on z.ai
-			expectedOutput:    0,
-			expectedCacheRead: 0,
+			expectedOutput:    floatPtr(0),
+			expectedCacheRead: floatPtr(0),
 		},
 		{
 			name:              "glm 4.7-flash free tier",
 			model:             "glm-4.7-flash",
 			expectedInput:     0,
-			expectedOutput:    0,
-			expectedCacheRead: 0,
+			expectedOutput:    floatPtr(0),
+			expectedCacheRead: floatPtr(0),
 		},
 		{
-			name:              "glm 4-32b legacy",
-			model:             "glm-4-32b-0414-128k",
-			expectedInput:     0.1e-6,
-			expectedOutput:    0.1e-6,
-			expectedCacheRead: 0,
+			name:           "glm 4-32b legacy",
+			model:          "glm-4-32b-0414-128k",
+			expectedInput:  0.1e-6,
+			expectedOutput: floatPtr(0.1e-6),
 		},
 		// 关键：5.1 必须先于 5 匹配（避免被 glm-5 抢走）
 		{
 			name:              "glm 5.1 vs glm 5 ordering (verbatim 5.1)",
 			model:             "glm-5.1",
 			expectedInput:     1.4e-6, // = glm-5.1 价格
-			expectedOutput:    4.4e-6,
-			expectedCacheRead: 0.26e-6,
+			expectedOutput:    floatPtr(4.4e-6),
+			expectedCacheRead: floatPtr(0.26e-6),
 		},
 		{
 			name:              "glm 4.5-air vs glm 4.5 ordering",
 			model:             "glm-4.5-air",
 			expectedInput:     0.2e-6, // = glm-4.5-air 价格（不是 glm-4.5 的 0.6e-6）
-			expectedOutput:    1.1e-6,
-			expectedCacheRead: 0.03e-6,
+			expectedOutput:    floatPtr(1.1e-6),
+			expectedCacheRead: floatPtr(0.03e-6),
 		},
 
 		// ---- 月之暗面 Kimi ----
@@ -490,44 +491,44 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			name:              "kimi k2.6 flagship",
 			model:             "kimi-k2.6",
 			expectedInput:     0.95e-6,
-			expectedOutput:    4e-6,
-			expectedCacheRead: 0.15e-6,
+			expectedOutput:    floatPtr(4e-6),
+			expectedCacheRead: floatPtr(0.15e-6),
 		},
 		{
 			name:              "kimi k2.5",
 			model:             "kimi-k2.5",
 			expectedInput:     0.60e-6,
-			expectedOutput:    3e-6,
-			expectedCacheRead: 0.098e-6,
+			expectedOutput:    floatPtr(3e-6),
+			expectedCacheRead: floatPtr(0.098e-6),
 		},
 		{
 			name:              "kimi k2-thinking",
 			model:             "kimi-k2-thinking",
 			expectedInput:     0.56e-6,
-			expectedOutput:    2.24e-6,
-			expectedCacheRead: 0.14e-6,
+			expectedOutput:    floatPtr(2.24e-6),
+			expectedCacheRead: floatPtr(0.14e-6),
 		},
 		{
 			name:              "kimi k2 base",
 			model:             "kimi-k2",
 			expectedInput:     0.56e-6,
-			expectedOutput:    2.24e-6,
-			expectedCacheRead: 0.14e-6,
+			expectedOutput:    floatPtr(2.24e-6),
+			expectedCacheRead: floatPtr(0.14e-6),
 		},
 		// 关键：k2.6 / k2.5 / k2-thinking 必须先于 k2 匹配
 		{
 			name:              "kimi k2.6 vs k2 ordering",
 			model:             "kimi-k2.6",
 			expectedInput:     0.95e-6, // = k2.6 不是 k2 的 0.56e-6
-			expectedOutput:    4e-6,
-			expectedCacheRead: 0.15e-6,
+			expectedOutput:    floatPtr(4e-6),
+			expectedCacheRead: floatPtr(0.15e-6),
 		},
 		{
 			name:              "kimi k2 thinking hyphenated variant",
 			model:             "kimi-k2-thinking-preview",
 			expectedInput:     0.56e-6,
-			expectedOutput:    2.24e-6,
-			expectedCacheRead: 0.14e-6,
+			expectedOutput:    floatPtr(2.24e-6),
+			expectedCacheRead: floatPtr(0.14e-6),
 		},
 
 		// ---- MiniMax M 系列 ----
@@ -535,46 +536,46 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			name:              "minimax m3",
 			model:             "minimax-m3",
 			expectedInput:     0.60e-6,
-			expectedOutput:    2.40e-6,
-			expectedCacheRead: 0.12e-6,
+			expectedOutput:    floatPtr(2.40e-6),
+			expectedCacheRead: floatPtr(0.12e-6),
 		},
 		{
 			name:              "minimax m3 long ctx boundary keep standard tier",
 			model:             "minimax-m3-long", // 仍按 standard tier (≤512K)
 			expectedInput:     0.60e-6,
-			expectedOutput:    2.40e-6,
-			expectedCacheRead: 0.12e-6,
+			expectedOutput:    floatPtr(2.40e-6),
+			expectedCacheRead: floatPtr(0.12e-6),
 		},
 		{
 			name:              "minimax m2.7",
 			model:             "minimax-m2.7",
 			expectedInput:     0.30e-6,
-			expectedOutput:    1.20e-6,
-			expectedCacheRead: 0.06e-6,
+			expectedOutput:    floatPtr(1.20e-6),
+			expectedCacheRead: floatPtr(0.06e-6),
 		},
 		{
 			name:              "minimax m2.7 highspeed",
 			model:             "minimax-m2.7-highspeed",
 			expectedInput:     0.60e-6,
-			expectedOutput:    2.40e-6,
-			expectedCacheRead: 0.06e-6,
+			expectedOutput:    floatPtr(2.40e-6),
+			expectedCacheRead: floatPtr(0.06e-6),
 		},
 		{
 			name:              "minimax m2.5",
 			model:             "minimax-m2.5",
 			expectedInput:     0.30e-6,
-			expectedOutput:    1.20e-6,
-			expectedCacheRead: 0.03e-6,
+			expectedOutput:    floatPtr(1.20e-6),
+			expectedCacheRead: floatPtr(0.03e-6),
 		},
 		{
 			name:              "minimax m2 legacy",
 			model:             "minimax-m2",
 			expectedInput:     0.30e-6,
-			expectedOutput:    1.20e-6,
-			expectedCacheRead: 0.03e-6,
+			expectedOutput:    floatPtr(1.20e-6),
+			expectedCacheRead: floatPtr(0.03e-6),
 		},
 
-		// ---- 负向用例：未覆盖的国产厂商 / alias 不应误计价 ----
+		// ---- 负向用例 ----
 		{name: "qwen unknown no fallback", model: "qwen-max", expectNilPricing: true},
 		{name: "doubao unknown no fallback", model: "doubao-pro", expectNilPricing: true},
 		{name: "hunyuan unknown no fallback", model: "hunyuan-t1", expectNilPricing: true},
@@ -585,8 +586,8 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			name:              "kimi k2-0905-preview implicit fallback to k2",
 			model:             "kimi-k2-0905-preview",
 			expectedInput:     0.56e-6,
-			expectedOutput:    2.24e-6,
-			expectedCacheRead: 0.14e-6,
+			expectedOutput:    floatPtr(2.24e-6),
+			expectedCacheRead: floatPtr(0.14e-6),
 		},
 	}
 
@@ -599,12 +600,12 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			}
 			require.NotNil(t, pricing)
 			require.InDelta(t, tt.expectedInput, pricing.InputPricePerToken, 1e-12)
-			if tt.expectedOutput != 0 {
-				require.InDelta(t, tt.expectedOutput, pricing.OutputPricePerToken, 1e-12,
+			if tt.expectedOutput != nil {
+				require.InDelta(t, *tt.expectedOutput, pricing.OutputPricePerToken, 1e-12,
 					"OutputPricePerToken mismatch for %s", tt.model)
 			}
-			if tt.expectedCacheRead != 0 {
-				require.InDelta(t, tt.expectedCacheRead, pricing.CacheReadPricePerToken, 1e-14,
+			if tt.expectedCacheRead != nil {
+				require.InDelta(t, *tt.expectedCacheRead, pricing.CacheReadPricePerToken, 1e-14,
 					"CacheReadPricePerToken mismatch for %s", tt.model)
 			}
 		})

From 4f5f2788e1f9b75c193c5eb05b2d229447bd3cdf Mon Sep 17 00:00:00 2001
From: alfadb <alfadb@163.com>
Date: Fri, 12 Jun 2026 13:38:19 +0800
Subject: [PATCH 7/8] fix(billing): add kimi-for-coding fallback pricing

---
 backend/internal/service/billing_service.go      | 12 +++++++++++-
 backend/internal/service/billing_service_test.go |  7 +++++++
 2 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/backend/internal/service/billing_service.go b/backend/internal/service/billing_service.go
index 43e91ce3e2..da36ebdaad 100644
--- a/backend/internal/service/billing_service.go
+++ b/backend/internal/service/billing_service.go
@@ -420,6 +420,13 @@ func (s *BillingService) initFallbackPricing() {
 		CacheReadPricePerToken: 0.15e-6, // $0.15 per MTok (cache hit, ¥1.10)
 		SupportsCacheBreakdown: false,
 	}
+	// kimi-for-coding 走 Kimi Coding endpoint，按当前 K2.6 coding 档位兜底计费。
+	s.fallbackPrices["kimi-for-coding"] = &ModelPricing{
+		InputPricePerToken:     0.95e-6,
+		OutputPricePerToken:    4e-6,
+		CacheReadPricePerToken: 0.15e-6,
+		SupportsCacheBreakdown: false,
+	}
 	s.fallbackPrices["kimi-k2.5"] = &ModelPricing{
 		InputPricePerToken:     0.60e-6, // $0.60 per MTok
 		OutputPricePerToken:    3e-6,    // $3.00 per MTok
@@ -576,8 +583,11 @@ func (s *BillingService) getFallbackPricing(model string) *ModelPricing {
 		return s.fallbackPrices["glm-4-32b-0414-128k"]
 	}
 
-	// 月之暗面 Kimi（kimi-k2.6 / kimi-k2.5 / kimi-k2-thinking / kimi-k2）
+	// 月之暗面 Kimi（kimi-k2.6 / kimi-for-coding / kimi-k2.5 / kimi-k2-thinking / kimi-k2）
 	// K2-0905 / K2-0711 官方未保留定价，不进入 fallback。
+	if strings.Contains(modelLower, "kimi-for-coding") {
+		return s.fallbackPrices["kimi-for-coding"]
+	}
 	if strings.Contains(modelLower, "kimi-k2.6") || strings.Contains(modelLower, "kimi-k2-6") {
 		return s.fallbackPrices["kimi-k2.6"]
 	}
diff --git a/backend/internal/service/billing_service_test.go b/backend/internal/service/billing_service_test.go
index 8bd6e41ede..45df3cd6bf 100644
--- a/backend/internal/service/billing_service_test.go
+++ b/backend/internal/service/billing_service_test.go
@@ -494,6 +494,13 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			expectedOutput:    floatPtr(4e-6),
 			expectedCacheRead: floatPtr(0.15e-6),
 		},
+		{
+			name:              "kimi for coding explicit alias",
+			model:             "kimi-for-coding",
+			expectedInput:     0.95e-6,
+			expectedOutput:    floatPtr(4e-6),
+			expectedCacheRead: floatPtr(0.15e-6),
+		},
 		{
 			name:              "kimi k2.5",
 			model:             "kimi-k2.5",

From 262fe1230d923f4f3e33a1ad3ace1e2f20a1693e Mon Sep 17 00:00:00 2001
From: alfadb <alfadb@163.com>
Date: Mon, 15 Jun 2026 12:02:06 +0800
Subject: [PATCH 8/8] =?UTF-8?q?feat(billing):=20=E4=B8=BA=20doubao-embeddi?=
 =?UTF-8?q?ng-vision=20=E6=B7=BB=E5=8A=A0=E5=9B=BE=E6=96=87=E5=B7=AE?=
 =?UTF-8?q?=E5=88=AB=E5=85=9C=E5=BA=95=E5=AE=9A=E4=BB=B7?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

火山方舟 doubao-embedding-vision 多模态向量化按量付费官方价为文本 ¥0.7/MTok、图片 ¥1.8/MTok，二者不同价。原计费引擎对 embedding 仅记单一 input token、无图片输入档位，无法表达该差别。

变更：
- ModelPricing 新增 ImageInputPricePerToken 字段
- OpenAIUsage / UsageTokens 新增 ImageInputTokens 字段
- extractOpenAIEmbeddingsUsage 解析 usage.prompt_tokens_details.image_tokens
- CalculateCost 拆分文本/图片输入计费；ImageInputTokens 为 0 时走原单价路径，存量 chat/vision 流量行为不变
- getFallbackPricing 新增 doubao-embedding-vision 分支（most-specific-first，覆盖带版本后缀别名），fallback 表填入 $0.098/$0.252 per MTok（汇率 ÷7.14）

测试：新增图文混合/纯文本/图片 token 超额三类用例及定价回退断言；service 包单测与 golangci-lint 全通过。
---
 backend/internal/service/billing_service.go   | 39 +++++++++++-
 .../internal/service/billing_service_test.go  | 61 +++++++++++++++++++
 backend/internal/service/openai_embeddings.go |  7 +++
 .../service/openai_gateway_service.go         |  2 +
 4 files changed, 108 insertions(+), 1 deletion(-)

diff --git a/backend/internal/service/billing_service.go b/backend/internal/service/billing_service.go
index da36ebdaad..74055a8151 100644
--- a/backend/internal/service/billing_service.go
+++ b/backend/internal/service/billing_service.go
@@ -91,6 +91,7 @@ type BillingCache interface {
 type ModelPricing struct {
 	InputPricePerToken             float64 // 每token输入价格 (USD)
 	InputPricePerTokenPriority     float64 // priority service tier 下每token输入价格 (USD)
+	ImageInputPricePerToken        float64 // 图片输入 token 价格 (USD)，用于多模态 embedding 等图文不同价场景；为 0 时回退到 InputPricePerToken
 	OutputPricePerToken            float64 // 每token输出价格 (USD)
 	OutputPricePerTokenPriority    float64 // priority service tier 下每token输出价格 (USD)
 	CacheCreationPricePerToken     float64 // 缓存创建每token价格 (USD)
@@ -137,6 +138,7 @@ func serviceTierCostMultiplier(serviceTier string) float64 {
 // UsageTokens 使用的token数量
 type UsageTokens struct {
 	InputTokens           int
+	ImageInputTokens      int
 	OutputTokens          int
 	CacheCreationTokens   int
 	CacheReadTokens       int
@@ -486,6 +488,17 @@ func (s *BillingService) initFallbackPricing() {
 		CacheReadPricePerToken: 0.03e-6,
 		SupportsCacheBreakdown: false,
 	}
+
+	// ---- 火山方舟 豆包 Embedding（多模态向量化）----
+	// doubao-embedding-vision 图文向量化：上游 usage 回传 prompt_tokens_details.{text_tokens,image_tokens}，
+	// 按量付费官方价 文本 ¥0.7/MTok、图片 ¥1.8/MTok；汇率口径 ÷7.14（与本表其他国产模型一致，¥1≈$0.14）。
+	// embedding 无 output，OutputPricePerToken 置 0。
+	s.fallbackPrices["doubao-embedding-vision"] = &ModelPricing{
+		InputPricePerToken:      0.098e-6, // ¥0.7/MTok ≈ $0.098（文本输入）
+		ImageInputPricePerToken: 0.252e-6, // ¥1.8/MTok ≈ $0.252（图片输入）
+		OutputPricePerToken:     0,
+		SupportsCacheBreakdown:  false,
+	}
 }
 
 // getFallbackPricing 根据模型系列获取回退价格
@@ -621,6 +634,13 @@ func (s *BillingService) getFallbackPricing(model string) *ModelPricing {
 		return s.fallbackPrices["minimax-m2"]
 	}
 
+	// 火山方舟 豆包 Embedding（多模态向量化）。
+	// most-specific-first：放在未来任何 doubao-embedding / doubao 宽匹配之前。
+	// 覆盖带版本后缀的别名（如 doubao-embedding-vision-251215）。
+	if strings.Contains(modelLower, "doubao-embedding-vision") {
+		return s.fallbackPrices["doubao-embedding-vision"]
+	}
+
 	// OpenAI（GPT-5 / Codex 族）：仅匹配已知型号，避免未知 OpenAI 型号误计价。
 	if normalized := normalizeKnownOpenAICodexModel(modelLower); normalized != "" {
 		switch normalized {
@@ -839,7 +859,24 @@ func (s *BillingService) computeTokenBreakdown(
 	}
 
 	bd := &CostBreakdown{}
-	bd.InputCost = float64(tokens.InputTokens) * inputPrice
+	// 分离图片输入 token 与文本输入 token（多模态 embedding 等图文不同价场景）。
+	// ImageInputTokens 为 0 时（绝大多数 chat/vision 流量）走原始单价路径，行为不变。
+	if tokens.ImageInputTokens > 0 {
+		imageInputTokens := tokens.ImageInputTokens
+		textInputTokens := tokens.InputTokens - imageInputTokens
+		if textInputTokens < 0 {
+			textInputTokens = 0
+			imageInputTokens = tokens.InputTokens
+		}
+		imageInputPrice := pricing.ImageInputPricePerToken
+		if imageInputPrice == 0 {
+			// 未配置图片输入档时回退到文本 input 价（已含 priority / 长上下文调整）
+			imageInputPrice = inputPrice
+		}
+		bd.InputCost = float64(textInputTokens)*inputPrice + float64(imageInputTokens)*imageInputPrice
+	} else {
+		bd.InputCost = float64(tokens.InputTokens) * inputPrice
+	}
 
 	// 分离图片输出 token 与文本输出 token
 	textOutputTokens := tokens.OutputTokens - tokens.ImageOutputTokens
diff --git a/backend/internal/service/billing_service_test.go b/backend/internal/service/billing_service_test.go
index 45df3cd6bf..9541047d25 100644
--- a/backend/internal/service/billing_service_test.go
+++ b/backend/internal/service/billing_service_test.go
@@ -582,9 +582,24 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			expectedCacheRead: floatPtr(0.03e-6),
 		},
 
+		// ---- 火山方舟 豆包 Embedding（多模态向量化）----
+		{
+			name:           "doubao embedding vision text rate",
+			model:          "doubao-embedding-vision",
+			expectedInput:  0.098e-6,
+			expectedOutput: floatPtr(0),
+		},
+		{
+			name:          "doubao embedding vision versioned alias",
+			model:         "doubao-embedding-vision-251215",
+			expectedInput: 0.098e-6,
+		},
+
 		// ---- 负向用例 ----
 		{name: "qwen unknown no fallback", model: "qwen-max", expectNilPricing: true},
+		// doubao-pro / doubao-embedding（纯文本）不在白名单，不回退；仅 doubao-embedding-vision 显式命中。
 		{name: "doubao unknown no fallback", model: "doubao-pro", expectNilPricing: true},
+		{name: "doubao text embedding no fallback", model: "doubao-embedding-text-240515", expectNilPricing: true},
 		{name: "hunyuan unknown no fallback", model: "hunyuan-t1", expectNilPricing: true},
 		{name: "moonshot v1 not covered", model: "moonshot-v1-8k", expectNilPricing: true},
 		// kimi-k2-0905 / kimi-k2-0711 官方未公布独立价，走 kimi-k2 隐性回退（接受）——
@@ -618,6 +633,52 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 		})
 	}
 }
+
+// doubao-embedding-vision 是首个图文不同价的 embedding：文本 ¥0.7/MTok、图片 ¥1.8/MTok。
+// 验证回退表同时携带文本与图片两档单价，且能被带版本后缀 / 大小写别名命中。
+func TestGetModelPricing_DoubaoEmbeddingVisionImageInputRate(t *testing.T) {
+	svc := newTestBillingService()
+
+	for _, model := range []string{
+		"doubao-embedding-vision",
+		"doubao-embedding-vision-251215",
+		"Doubao-Embedding-Vision",
+	} {
+		pricing, err := svc.GetModelPricing(model)
+		require.NoError(t, err, "model %s should resolve fallback pricing", model)
+		require.NotNil(t, pricing)
+		require.InDelta(t, 0.098e-6, pricing.InputPricePerToken, 1e-12, "text input rate for %s", model)
+		require.InDelta(t, 0.252e-6, pricing.ImageInputPricePerToken, 1e-12, "image input rate for %s", model)
+		require.Zero(t, pricing.OutputPricePerToken, "embedding has no output cost for %s", model)
+	}
+}
+
+// 验证双档计费：InputCost = 文本token×文本价 + 图片token×图片价；
+// 且 ImageInputTokens=0 时走原单价路径，ImageInputTokens>InputTokens 时不负计文本。
+func TestCalculateCost_DoubaoEmbeddingVisionDifferentialInput(t *testing.T) {
+	svc := newTestBillingService()
+
+	// 图文混合：prompt_tokens=1340，其中 image_tokens=28、text_tokens=1312。
+	mixed := UsageTokens{InputTokens: 1340, ImageInputTokens: 28}
+	cost, err := svc.CalculateCost("doubao-embedding-vision", mixed, 1.0)
+	require.NoError(t, err)
+	wantMixed := float64(1312)*0.098e-6 + float64(28)*0.252e-6
+	require.InDelta(t, wantMixed, cost.InputCost, 1e-15)
+	require.InDelta(t, wantMixed, cost.TotalCost, 1e-15)
+	require.Zero(t, cost.OutputCost)
+
+	// 纯文本：全部按文本档计费，与原单价路径一致。
+	textOnly := UsageTokens{InputTokens: 1340}
+	costText, err := svc.CalculateCost("doubao-embedding-vision", textOnly, 1.0)
+	require.NoError(t, err)
+	require.InDelta(t, float64(1340)*0.098e-6, costText.InputCost, 1e-15)
+
+	// 健壮性：ImageInputTokens 超过 InputTokens 时，文本置 0、计费 token 不超过 InputTokens。
+	weird := UsageTokens{InputTokens: 10, ImageInputTokens: 50}
+	costWeird, err := svc.CalculateCost("doubao-embedding-vision", weird, 1.0)
+	require.NoError(t, err)
+	require.InDelta(t, float64(10)*0.252e-6, costWeird.InputCost, 1e-15)
+}
 func TestCalculateCostWithLongContext_BelowThreshold(t *testing.T) {
 	svc := newTestBillingService()
 
diff --git a/backend/internal/service/openai_embeddings.go b/backend/internal/service/openai_embeddings.go
index 7c710259f3..0fb3fff1f7 100644
--- a/backend/internal/service/openai_embeddings.go
+++ b/backend/internal/service/openai_embeddings.go
@@ -214,8 +214,15 @@ func extractOpenAIEmbeddingsUsage(body []byte) OpenAIUsage {
 		usage.Get("cache_creation_input_tokens"),
 		usage.Get("input_tokens_details.cache_creation_tokens"),
 	)
+	// 多模态 embedding（如 doubao-embedding-vision）回传图文 token 拆分，
+	// 用于图文不同价计费；纯文本 embedding 该字段为 0，行为不变。
+	imageInputTokens := firstPositiveGJSONInt(
+		usage.Get("prompt_tokens_details.image_tokens"),
+		usage.Get("input_tokens_details.image_tokens"),
+	)
 	return OpenAIUsage{
 		InputTokens:              inputTokens,
+		ImageInputTokens:         imageInputTokens,
 		OutputTokens:             outputTokens,
 		CacheReadInputTokens:     cacheReadTokens,
 		CacheCreationInputTokens: cacheCreationTokens,
diff --git a/backend/internal/service/openai_gateway_service.go b/backend/internal/service/openai_gateway_service.go
index d056668a48..7fb53e2f55 100644
--- a/backend/internal/service/openai_gateway_service.go
+++ b/backend/internal/service/openai_gateway_service.go
@@ -212,6 +212,7 @@ func (s *OpenAICodexUsageSnapshot) Normalize() *NormalizedCodexLimits {
 // OpenAIUsage represents OpenAI API response usage
 type OpenAIUsage struct {
 	InputTokens              int `json:"input_tokens"`
+	ImageInputTokens         int `json:"image_input_tokens,omitempty"`
 	OutputTokens             int `json:"output_tokens"`
 	CacheCreationInputTokens int `json:"cache_creation_input_tokens,omitempty"`
 	CacheReadInputTokens     int `json:"cache_read_input_tokens,omitempty"`
@@ -5912,6 +5913,7 @@ func (s *OpenAIGatewayService) RecordUsage(ctx context.Context, input *OpenAIRec
 	// Calculate cost
 	tokens := UsageTokens{
 		InputTokens:         actualInputTokens,
+		ImageInputTokens:    result.Usage.ImageInputTokens,
 		OutputTokens:        result.Usage.OutputTokens,
 		CacheCreationTokens: result.Usage.CacheCreationInputTokens,
 		CacheReadTokens:     result.Usage.CacheReadInputTokens,