Merge pull request #3243 from alfadb/feature/chinese-llm-fallback-pricing

feat(billing): 国产 LLM 兜底定价 (GLM / Kimi / MiniMax) + 收编 DeepSeek V4

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
@@ -305,6 +307,198 @@ 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,
+	}
+	// 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
+		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,
+	}
+
+	// ---- 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
@@ -344,7 +538,110 @@ func (s *BillingService) getFallbackPricing(model string) *ModelPricing {
 		return s.fallbackPrices["gemini-3.1-pro"]
 	}
 
-	// OpenAI 仅匹配已知 GPT-5/Codex 族，避免未知 OpenAI 型号误计价。
+	// 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"]
+	}
+	if strings.Contains(modelLower, "deepseek-chat") || strings.Contains(modelLower, "deepseek-reasoner") {
+		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-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"]
+	}
+	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") {
+		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"]
+	}
+
+	// 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 {
 		case "gpt-5.5":
@@ -562,7 +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

backend/internal/service/billing_service_test.go

@@ -332,11 +332,16 @@ func TestCalculateCost_LongContextAppliesMultiplierToCacheCreation5mAnd1h(t *tes
 func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 	svc := newTestBillingService()
 
+	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
-		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},
@@ -352,7 +357,260 @@ 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: "non supported family", model: "qwen-max", expectNilPricing: true},
+		{
+			name:              "deepseek v4 pro",
+			model:             "deepseek-v4-pro",
+			expectedInput:     4.35e-7,
+			expectedOutput:    floatPtr(8.7e-7),
+			expectedCacheRead: floatPtr(3.625e-9),
+		},
+		{
+			name:              "deepseek v4 flash",
+			model:             "deepseek-v4-flash",
+			expectedInput:     1.4e-7,
+			expectedOutput:    floatPtr(2.8e-7),
+			expectedCacheRead: floatPtr(2.8e-9),
+		},
+		{
+			name:              "deepseek chat alias → flash",
+			model:             "deepseek-chat",
+			expectedInput:     1.4e-7,
+			expectedOutput:    floatPtr(2.8e-7),
+			expectedCacheRead: floatPtr(2.8e-9),
+		},
+		{
+			name:              "deepseek reasoner alias → flash",
+			model:             "deepseek-reasoner",
+			expectedInput:     1.4e-7,
+			expectedOutput:    floatPtr(2.8e-7),
+			expectedCacheRead: floatPtr(2.8e-9),
+		},
+
+		// ---- 智谱 GLM（z.ai USD 口径）----
+		{
+			name:              "glm 5.1 flagship",
+			model:             "glm-5.1",
+			expectedInput:     1.4e-6,
+			expectedOutput:    floatPtr(4.4e-6),
+			expectedCacheRead: floatPtr(0.26e-6),
+		},
+		{
+			name:              "glm 5 base",
+			model:             "glm-5",
+			expectedInput:     1e-6,
+			expectedOutput:    floatPtr(3.2e-6),
+			expectedCacheRead: floatPtr(0.2e-6),
+		},
+		{
+			name:              "glm 5 turbo",
+			model:             "glm-5-turbo",
+			expectedInput:     1.2e-6,
+			expectedOutput:    floatPtr(4e-6),
+			expectedCacheRead: floatPtr(0.24e-6),
+		},
+		{
+			name:              "glm 4.7",
+			model:             "glm-4.7",
+			expectedInput:     0.6e-6,
+			expectedOutput:    floatPtr(2.2e-6),
+			expectedCacheRead: floatPtr(0.11e-6),
+		},
+		{
+			name:              "glm 4.6",
+			model:             "glm-4.6",
+			expectedInput:     0.6e-6,
+			expectedOutput:    floatPtr(2.2e-6),
+			expectedCacheRead: floatPtr(0.11e-6),
+		},
+		{
+			name:              "glm 4.5",
+			model:             "glm-4.5",
+			expectedInput:     0.6e-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:    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:    floatPtr(1.1e-6),
+			expectedCacheRead: floatPtr(0.03e-6),
+		},
+		{
+			name:              "glm 4.7-flashx",
+			model:             "glm-4.7-flashx",
+			expectedInput:     0.07e-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:    floatPtr(0),
+			expectedCacheRead: floatPtr(0),
+		},
+		{
+			name:              "glm 4.7-flash free tier",
+			model:             "glm-4.7-flash",
+			expectedInput:     0,
+			expectedOutput:    floatPtr(0),
+			expectedCacheRead: floatPtr(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:    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:    floatPtr(1.1e-6),
+			expectedCacheRead: floatPtr(0.03e-6),
+		},
+
+		// ---- 月之暗面 Kimi ----
+		{
+			name:              "kimi k2.6 flagship",
+			model:             "kimi-k2.6",
+			expectedInput:     0.95e-6,
+			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",
+			expectedInput:     0.60e-6,
+			expectedOutput:    floatPtr(3e-6),
+			expectedCacheRead: floatPtr(0.098e-6),
+		},
+		{
+			name:              "kimi k2-thinking",
+			model:             "kimi-k2-thinking",
+			expectedInput:     0.56e-6,
+			expectedOutput:    floatPtr(2.24e-6),
+			expectedCacheRead: floatPtr(0.14e-6),
+		},
+		{
+			name:              "kimi k2 base",
+			model:             "kimi-k2",
+			expectedInput:     0.56e-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:    floatPtr(4e-6),
+			expectedCacheRead: floatPtr(0.15e-6),
+		},
+		{
+			name:              "kimi k2 thinking hyphenated variant",
+			model:             "kimi-k2-thinking-preview",
+			expectedInput:     0.56e-6,
+			expectedOutput:    floatPtr(2.24e-6),
+			expectedCacheRead: floatPtr(0.14e-6),
+		},
+
+		// ---- MiniMax M 系列 ----
+		{
+			name:              "minimax m3",
+			model:             "minimax-m3",
+			expectedInput:     0.60e-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:    floatPtr(2.40e-6),
+			expectedCacheRead: floatPtr(0.12e-6),
+		},
+		{
+			name:              "minimax m2.7",
+			model:             "minimax-m2.7",
+			expectedInput:     0.30e-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:    floatPtr(2.40e-6),
+			expectedCacheRead: floatPtr(0.06e-6),
+		},
+		{
+			name:              "minimax m2.5",
+			model:             "minimax-m2.5",
+			expectedInput:     0.30e-6,
+			expectedOutput:    floatPtr(1.20e-6),
+			expectedCacheRead: floatPtr(0.03e-6),
+		},
+		{
+			name:              "minimax m2 legacy",
+			model:             "minimax-m2",
+			expectedInput:     0.30e-6,
+			expectedOutput:    floatPtr(1.20e-6),
+			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）——
+		// 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:    floatPtr(2.24e-6),
+			expectedCacheRead: floatPtr(0.14e-6),
+		},
 	}
 
 	for _, tt := range tests {
@@ -364,9 +622,63 @@ func TestGetFallbackPricing_FamilyMatching(t *testing.T) {
 			}
 			require.NotNil(t, pricing)
 			require.InDelta(t, tt.expectedInput, pricing.InputPricePerToken, 1e-12)
+			if tt.expectedOutput != nil {
+				require.InDelta(t, *tt.expectedOutput, pricing.OutputPricePerToken, 1e-12,
+					"OutputPricePerToken mismatch for %s", tt.model)
+			}
+			if tt.expectedCacheRead != nil {
+				require.InDelta(t, *tt.expectedCacheRead, pricing.CacheReadPricePerToken, 1e-14,
+					"CacheReadPricePerToken mismatch for %s", tt.model)
+			}
 		})
 	}
 }
+
+// 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()
 

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,

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)

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,