refactor(billing): 长上下文阶梯计价改为价格目录数据驱动

删除 GPT/Gemini 系列写死的阶梯规则，阶梯一律由目录数据驱动： - 解析层折算 LiteLLM 的 *_above_XXXk_tokens 绝对价字段为阈值+倍率（阈值取自字段名，倍率 = above/基础价，cache 侧跟随输入倍率， xAI 按 provider 用达到即进高档语义）；显式 long_context_* 字段按"字段存在"优先，显式 threshold=0 或 multiplier=1 可关闭阶梯 - 删除 OpenAI 272K 常量与 applyModelSpecificPricingPolicy 的长上下文强补分支、fallbackPrices/静态兜底条目的 LongContext 字段、Gemini 旧边际规则全链路（CalculateCostWithLongContext / RecordUsageWithLongContext / 阶梯表 marginal 基准）；保留 fastRatio 与 GPT-5.6 cache_write 1.25x 防御分支、grok 静态兜底阶梯 - 渠道配置定价区间时完全以渠道区间为准，不再叠加目录阶梯 - 长上下文倍率 ≤0 在应用点按 1 计，防止显式字段残缺把分项乘 0 免费 - 目录重载时对比新旧数据，原有阶梯丢失打 WARN 哨兵告警行为变化：Gemini 由边际 2x（仅 /v1beta、全模型）改为整单换档（数据驱动、仅带 above_200k 的 pro 系），渠道平价之上叠加目录阶梯；Claude sonnet-4/4.5 随目录数据获得 200K 整单阶梯（与官方定价一致）；模型广场官方参考价列与实付列同源展示目录阶梯。
2026-10-19 16:17:56 +00:00 · 2026-09-01 01:26:55 +08:00
parent 200602b41b
commit 530fb20f25
17 changed files with 599 additions and 633 deletions
@@ -2279,7 +2279,7 @@ func setDefaults() {
 	viper.SetDefault("rate_limit.overload_cooldown_minutes", 10)
 	viper.SetDefault("rate_limit.oauth_401_cooldown_minutes", 10)

-	// Pricing - 从 model-price-repo 同步模型定价和上下文窗口数据（固定到 commit，避免分支漂移）
+	// Pricing - 从 model-price-repo main 分支同步模型定价和上下文窗口数据
 	viper.SetDefault("pricing.remote_url", "https://raw.githubusercontent.com/Wei-Shaw/model-price-repo/main/model_prices_and_context_window.json")
 	viper.SetDefault("pricing.hash_url", "https://raw.githubusercontent.com/Wei-Shaw/model-price-repo/main/model_prices_and_context_window.sha256")
 	viper.SetDefault("pricing.data_dir", "./data")
@@ -279,7 +279,7 @@ func InboundEndpointMiddleware() gin.HandlerFunc {

 // ──────────────────────────────────────────────────────────
 // Context helpers — used by handlers before building
-// RecordUsageInput / RecordUsageLongContextInput.
+// RecordUsageInput.
 // ──────────────────────────────────────────────────────────

 // GetInboundEndpoint returns the canonical inbound endpoint stored by
@@ -569,33 +569,25 @@ func (h *GatewayHandler) GeminiV1BetaModels(c *gin.Context) {
 		forceCacheBilling := fs.ForceCacheBilling
 		quotaPlatform := service.QuotaPlatform(c.Request.Context(), apiKey)
 		sessionID := service.ExtractClientSessionID(c)
-		// 长上下文规则由计费服务统一持有（模型广场展示同源），入口只负责声明自己适用该规则。
-		var longContextThreshold int
-		var longContextMultiplier float64
-		if rule := h.gatewayService.LegacyLongContextRule(service.PlatformGemini); rule != nil {
-			longContextThreshold = rule.Threshold
-			longContextMultiplier = rule.Multiplier
-		}
+		// 长上下文阶梯由目录数据驱动，统一在计费路径内生效，入口无需声明。
 		h.submitUsageRecordTask(c.Request.Context(), func(ctx context.Context) {
-			if err := h.gatewayService.RecordUsageWithLongContext(ctx, &service.RecordUsageLongContextInput{
-				Result:                result,
-				QuotaPlatform:         quotaPlatform,
-				APIKey:                apiKey,
-				User:                  apiKey.User,
-				Account:               account,
-				Subscription:          subscription,
-				PricingAt:             pricingAt,
-				InboundEndpoint:       inboundEndpoint,
-				UpstreamEndpoint:      upstreamEndpoint,
-				UserAgent:             userAgent,
-				IPAddress:             clientIP,
-				RequestPayloadHash:    requestPayloadHash,
-				LongContextThreshold:  longContextThreshold,
-				LongContextMultiplier: longContextMultiplier,
-				ForceCacheBilling:     forceCacheBilling,
-				APIKeyService:         h.apiKeyService,
-				SessionID:             sessionID,
-				ChannelUsageFields:    clientRequestedUsageFields(c, channelMapping, reqModel, result.UpstreamModel),
+			if err := h.gatewayService.RecordUsage(ctx, &service.RecordUsageInput{
+				Result:             result,
+				QuotaPlatform:      quotaPlatform,
+				APIKey:             apiKey,
+				User:               apiKey.User,
+				Account:            account,
+				Subscription:       subscription,
+				PricingAt:          pricingAt,
+				InboundEndpoint:    inboundEndpoint,
+				UpstreamEndpoint:   upstreamEndpoint,
+				UserAgent:          userAgent,
+				IPAddress:          clientIP,
+				RequestPayloadHash: requestPayloadHash,
+				ForceCacheBilling:  forceCacheBilling,
+				APIKeyService:      h.apiKeyService,
+				SessionID:          sessionID,
+				ChannelUsageFields: clientRequestedUsageFields(c, channelMapping, reqModel, result.UpstreamModel),
 			}); err != nil {
 				logger.L().With(
 					zap.String("component", "handler.gemini_v1beta.models"),
@@ -9,15 +9,12 @@ import (
 	"time"
 )

-// ContextPricingBasis 阶梯的计价基准。
+// ContextPricingBasis 阶梯的计价基准。当前只有整单口径；历史上的
+// Gemini 边际口径（"marginal"）已随平台旧规则一并移除。
 type ContextPricingBasis string

-const (
-	// ContextPricingBasisWholeRequest 整单按所在档单价计价（目录阶梯、渠道区间）。
-	ContextPricingBasisWholeRequest ContextPricingBasis = "whole_request"
-	// ContextPricingBasisMarginal 仅超出阈值的部分按该档单价计价（平台旧规则）。
-	ContextPricingBasisMarginal ContextPricingBasis = "marginal"
-)
+// ContextPricingBasisWholeRequest 整单按所在档单价计价（目录阶梯、渠道区间）。
+const ContextPricingBasisWholeRequest ContextPricingBasis = "whole_request"

 // ContextPricingTier (MinTokens, MaxTokens] 区间内的有效 per-token 单价（USD）。
 // nil 表示该项无价/不计费；MaxTokens 为 nil 表示无上限。
@@ -73,8 +70,8 @@ const contextProbeDelta = 1000
 // ResolveContextPricingSchedule 解析分组+模型的上下文阶梯单价表。
 //
 // 解析链与扣费完全一致：Resolver.Resolve（分组卡 → 渠道 → 目录 → 策略）给出定价，
-// CalculateTokenCostForRequest 给出路径（分组/渠道定价 → 平台旧规则 → 内置目录）。
-// 断点只取自计费自身的规则输入（渠道区间边界、目录阶梯阈值、旧规则阈值），
+// CalculateTokenCostForRequest 给出路径（分组/渠道定价 → 内置目录）。
+// 断点只取自计费自身的规则输入（渠道区间边界、目录阶梯阈值），
 // 每一段的单价由真实计费函数在该段内两点探针的差商得到，因此倍率、策略等
 // 规则变更无需同步到这里；相邻同价段会合并。
 //
@@ -103,22 +100,13 @@ func (s *BillingService) ResolveContextPricingSchedule(ctx context.Context, reso
 		return nil, nil
 	}

-	var legacy *LegacyLongContextRule
-	if in.Group != nil {
-		legacy = s.LegacyLongContextRule(in.Platform)
-	}
-	if !legacyLongContextApplies(resolved, in.Group, legacy) {
-		legacy = nil
-	}
-
 	req := TokenCostRequest{
-		Ctx:               ctx,
-		Model:             in.Model,
-		Group:             in.Group,
-		RateMultiplier:    1,
-		Resolver:          resolver,
-		Resolved:          resolved,
-		LegacyLongContext: legacy,
+		Ctx:            ctx,
+		Model:          in.Model,
+		Group:          in.Group,
+		RateMultiplier: 1,
+		Resolver:       resolver,
+		Resolved:       resolved,
 	}
 	probe := func(tokens UsageTokens) (*CostBreakdown, error) {
 		r := req
@@ -126,7 +114,7 @@ func (s *BillingService) ResolveContextPricingSchedule(ctx context.Context, reso
 		return s.CalculateTokenCostForRequest(r)
 	}

-	plan := s.contextPricingBreakpoints(resolver, resolved, in.Model, legacy)
+	plan := s.contextPricingBreakpoints(resolver, resolved, in.Model)
 	segments := buildContextSegments(plan.bounds)

 	tiers := make([]ContextPricingTier, 0, len(segments))
@@ -140,11 +128,7 @@ func (s *BillingService) ResolveContextPricingSchedule(ctx context.Context, reso
 	tiers = mergeEqualContextTiers(tiers)
 	applyContextTierLabels(tiers, plan)

-	basis := ContextPricingBasisWholeRequest
-	if legacy != nil {
-		basis = ContextPricingBasisMarginal
-	}
-	return &ContextPricingSchedule{Basis: basis, Tiers: tiers, TimePricing: resolvedTimePricingSchedule(resolved)}, nil
+	return &ContextPricingSchedule{Basis: ContextPricingBasisWholeRequest, Tiers: tiers, TimePricing: resolvedTimePricingSchedule(resolved)}, nil
 }

 // resolvedTimePricingSchedule 列出计费会生效的分时倍率时段。
@@ -210,14 +194,8 @@ type contextBreakpointPlan struct {
 }

 // contextPricingBreakpoints 从计费自身的规则输入收集价格断点（不读取任何倍率）。
-func (s *BillingService) contextPricingBreakpoints(resolver *ModelPricingResolver, resolved *ResolvedPricing, model string, legacy *LegacyLongContextRule) contextBreakpointPlan {
+func (s *BillingService) contextPricingBreakpoints(resolver *ModelPricingResolver, resolved *ResolvedPricing, model string) contextBreakpointPlan {
 	plan := contextBreakpointPlan{}
-	if legacy != nil {
-		plan.bounds = []int{legacy.Threshold}
-		plan.thresholdBound = legacy.Threshold
-		plan.threshold = legacy.Threshold
-		return plan
-	}
 	if !resolved.longContextPricingEnabled {
 		return plan
 	}
@@ -69,12 +69,25 @@ func requirePrice(t *testing.T, want, got *float64, field string) {
 	require.InDelta(t, *want, *got, 1e-15, field)
 }

+// mustCatalogFromJSON 走生产解析路径（含 above_XXXk 折算）构造目录 stub；场景表无 *testing.T，解析失败直接 panic。
+func mustCatalogFromJSON(body string) *PricingService {
+	s := &PricingService{}
+	data, err := s.parsePricingData([]byte(body))
+	if err != nil {
+		panic(err)
+	}
+	s.pricingData = data
+	return s
+}
+
+func openAILadderCatalog() *PricingService { return mustCatalogFromJSON(openAILadderCatalogJSON) }
+
 func scheduleScenarios() []scheduleScenario {
 	p := testPtrFloat64
 	return []scheduleScenario{
 		{
 			name: "官方阶梯 gpt-5.4 整单两档", model: "gpt-5.4", platform: PlatformOpenAI, groupPlatform: PlatformOpenAI,
-			group: enabledGroup(PlatformOpenAI), wantBasis: ContextPricingBasisWholeRequest,
+			group: enabledGroup(PlatformOpenAI), catalog: openAILadderCatalog(), wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 2)
 				requireTier(t, s.Tiers[0], 0, intPtr(272000), "≤272K", p(2.5e-6), p(15e-6), p(2.5e-6), p(0.25e-6))
@@ -92,7 +105,7 @@ func scheduleScenarios() []scheduleScenario {
 		},
 		{
 			name: "分组关闭阶梯只剩基础档", model: "gpt-5.4", platform: PlatformOpenAI, groupPlatform: PlatformOpenAI,
-			group: disabledGroup(PlatformOpenAI), wantBasis: ContextPricingBasisWholeRequest,
+			group: disabledGroup(PlatformOpenAI), catalog: openAILadderCatalog(), wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 1)
 				requireTier(t, s.Tiers[0], 0, nil, "", p(2.5e-6), p(15e-6), p(2.5e-6), p(0.25e-6))
@@ -100,7 +113,7 @@ func scheduleScenarios() []scheduleScenario {
 		},
 		{
 			name: "官方参考价（无分组）带目录阶梯", model: "gpt-5.4", platform: "", groupPlatform: PlatformOpenAI,
-			group: nil, wantBasis: ContextPricingBasisWholeRequest,
+			group: nil, catalog: openAILadderCatalog(), wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 2)
 				requireTier(t, s.Tiers[1], 272000, nil, ">272K", p(5e-6), p(22.5e-6), p(5e-6), p(0.5e-6))
@@ -210,6 +223,7 @@ func scheduleScenarios() []scheduleScenario {
 			group: &Group{ID: 100, Platform: PlatformOpenAI, LongContextPricingEnabled: true, ModelPricing: []ChannelModelPricing{{
 				Models: []string{"gpt-5.4"}, BillingMode: BillingModeToken, InputPrice: p(1e-6),
 			}}},
+			catalog:   openAILadderCatalog(),
 			wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 2)
@@ -219,46 +233,55 @@ func scheduleScenarios() []scheduleScenario {
 			},
 		},
 		{
-			name: "Gemini 旧规则按超出部分计价", model: "gemini-2.5-pro", platform: PlatformGemini, groupPlatform: PlatformGemini,
-			group: enabledGroup(PlatformGemini), catalog: geminiCatalogStub(), wantBasis: ContextPricingBasisMarginal,
+			name: "Gemini 目录阶梯整单换档", model: "gemini-2.5-pro", platform: PlatformGemini, groupPlatform: PlatformGemini,
+			group: enabledGroup(PlatformGemini), catalog: mustCatalogFromJSON(geminiLadderCatalogJSON), wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 2)
 				requireTier(t, s.Tiers[0], 0, intPtr(200000), "≤200K", p(1.25e-6), p(10e-6), nil, p(0.3125e-6))
-				requireTier(t, s.Tiers[1], 200000, nil, ">200K", p(2.5e-6), p(10e-6), nil, p(0.625e-6))
+				requireTier(t, s.Tiers[1], 200000, nil, ">200K", p(2.5e-6), p(15e-6), nil, p(0.625e-6))
 			},
 		},
 		{
-			name: "Gemini 分组关闭时不用旧规则", model: "gemini-2.5-pro", platform: PlatformGemini, groupPlatform: PlatformGemini,
-			group: disabledGroup(PlatformGemini), catalog: geminiCatalogStub(), wantBasis: ContextPricingBasisWholeRequest,
+			name: "Gemini 分组关闭时无阶梯", model: "gemini-2.5-pro", platform: PlatformGemini, groupPlatform: PlatformGemini,
+			group: disabledGroup(PlatformGemini), catalog: mustCatalogFromJSON(geminiLadderCatalogJSON), wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 1)
 			},
 		},
 		{
-			name: "Gemini 有渠道定价时旧规则让位", model: "gemini-2.5-pro", platform: PlatformGemini, groupPlatform: PlatformGemini,
-			group: enabledGroup(PlatformGemini), catalog: geminiCatalogStub(),
+			name: "Gemini 渠道平价之上叠加目录阶梯", model: "gemini-2.5-pro", platform: PlatformGemini, groupPlatform: PlatformGemini,
+			group: enabledGroup(PlatformGemini), catalog: mustCatalogFromJSON(geminiLadderCatalogJSON),
 			channel: []ChannelModelPricing{{
 				Platform: PlatformGemini, Models: []string{"gemini-2.5-pro"}, BillingMode: BillingModeToken, InputPrice: p(3e-6),
 			}},
 			wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
-				require.Len(t, s.Tiers, 1)
+				require.Len(t, s.Tiers, 2)
 				requirePrice(t, p(3e-6), s.Tiers[0].Input, "input")
+				requirePrice(t, p(6e-6), s.Tiers[1].Input, "input")
 			},
 		},
 		{
-			name: "Gemini 官方参考不套用站内旧规则", model: "gemini-2.5-pro", platform: "", groupPlatform: PlatformGemini,
-			group: nil, catalog: geminiCatalogStub(), wantBasis: ContextPricingBasisWholeRequest,
+			name: "Gemini 目录无阶梯字段时开关开启也无阶梯", model: "gemini-2.5-pro", platform: PlatformGemini, groupPlatform: PlatformGemini,
+			group: enabledGroup(PlatformGemini), catalog: geminiCatalogStub(), wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 1)
 			},
 		},
+		{
+			name: "Gemini 官方参考价与目录数据同源", model: "gemini-2.5-pro", platform: "", groupPlatform: PlatformGemini,
+			group: nil, catalog: mustCatalogFromJSON(geminiLadderCatalogJSON), wantBasis: ContextPricingBasisWholeRequest,
+			check: func(t *testing.T, s *ContextPricingSchedule) {
+				require.Len(t, s.Tiers, 2)
+			},
+		},
 		{
 			name: "composite 分组按模型平台取渠道定价并叠加官方阶梯", model: "gpt-5.4", platform: PlatformOpenAI, groupPlatform: PlatformComposite,
 			group: enabledGroup(PlatformComposite),
 			channel: []ChannelModelPricing{{
 				Platform: PlatformOpenAI, Models: []string{"gpt-5.4"}, BillingMode: BillingModeToken, InputPrice: p(1e-6),
 			}},
+			catalog:   openAILadderCatalog(),
 			wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 2)
@@ -283,9 +306,11 @@ func scheduleScenarios() []scheduleScenario {
 		{
 			name: "gpt-5.6 缺 cache_write 时按策略补 1.25 倍并带阶梯", model: "gpt-5.6-sol", platform: PlatformOpenAI, groupPlatform: PlatformOpenAI,
 			group: enabledGroup(PlatformOpenAI),
-			catalog: newStubPricingServiceFromMap(map[string]*LiteLLMModelPricing{
-				"gpt-5.6-sol": {Mode: "chat", InputCostPerToken: 5e-6, OutputCostPerToken: 30e-6, CacheReadInputTokenCost: 0.5e-6},
-			}),
+			catalog: mustCatalogFromJSON(`{"gpt-5.6-sol": {"litellm_provider": "openai", "mode": "chat",
+				"input_cost_per_token": 5e-06, "output_cost_per_token": 3e-05, "cache_read_input_token_cost": 5e-07,
+				"input_cost_per_token_above_272k_tokens": 1e-05,
+				"output_cost_per_token_above_272k_tokens": 4.5e-05,
+				"cache_read_input_token_cost_above_272k_tokens": 1e-06}}`),
 			wantBasis: ContextPricingBasisWholeRequest,
 			check: func(t *testing.T, s *ContextPricingSchedule) {
 				require.Len(t, s.Tiers, 2)
@@ -398,22 +423,8 @@ func tierAt(tiers []ContextPricingTier, contextTokens int) ContextPricingTier {
 }

 // expectedCostFromSchedule 按阶梯表推算 contextTokens 个某类 token 的费用：
-// 整单基准取所在档单价 × 全量；边际基准逐段累加。
+// 整单基准取所在档单价 × 全量。
 func expectedCostFromSchedule(s *ContextPricingSchedule, kind tokenKind, contextTokens int) float64 {
-	if s.Basis == ContextPricingBasisMarginal {
-		total := 0.0
-		for _, tier := range s.Tiers {
-			if contextTokens <= tier.MinTokens {
-				break
-			}
-			upper := contextTokens
-			if tier.MaxTokens != nil && *tier.MaxTokens < upper {
-				upper = *tier.MaxTokens
-			}
-			total += float64(upper-tier.MinTokens) * tierPrice(tier, kind)
-		}
-		return total
-	}
 	return float64(contextTokens) * tierPrice(tierAt(s.Tiers, contextTokens), kind)
 }

@@ -443,17 +454,10 @@ func TestResolveContextPricingSchedule_ParityWithBilling(t *testing.T) {
 				pricingInput.GroupID = &gid
 			}
 			resolved := resolver.Resolve(ctx, pricingInput)
-			var legacy *LegacyLongContextRule
-			if sc.group != nil {
-				legacy = bs.LegacyLongContextRule(sc.platform)
-			}
-			if !legacyLongContextApplies(resolved, sc.group, legacy) {
-				legacy = nil
-			}
 			cost := func(tokens UsageTokens) float64 {
 				bd, err := bs.CalculateTokenCostForRequest(TokenCostRequest{
 					Ctx: ctx, Model: sc.model, Group: sc.group, Tokens: tokens, RateMultiplier: 1,
-					Resolver: resolver, Resolved: resolved, LegacyLongContext: legacy,
+					Resolver: resolver, Resolved: resolved,
 				})
 				require.NoError(t, err)
 				return bd.ActualCost
@@ -116,12 +116,6 @@ type ModelPricing struct {
 	ImageOutputPriceExplicit           bool     // 是否由渠道定价显式设定（为 true 时即使 == 0 也不回退）
 }

-const (
-	openAIGPT54LongContextInputThreshold   = 272000
-	openAIGPT54LongContextInputMultiplier  = 2.0
-	openAIGPT54LongContextOutputMultiplier = 1.5
-)
-
 func normalizeBillingServiceTier(serviceTier string) string {
 	return strings.ToLower(strings.TrimSpace(serviceTier))
 }
@@ -395,9 +389,6 @@ func (s *BillingService) initFallbackPricing() {
 		CacheReadPricePerToken:         0.25e-6, // $0.25 per MTok
 		CacheReadPricePerTokenPriority: 0.5e-6,  // $0.5 per MTok
 		SupportsCacheBreakdown:         false,
-		LongContextInputThreshold:      openAIGPT54LongContextInputThreshold,
-		LongContextInputMultiplier:     openAIGPT54LongContextInputMultiplier,
-		LongContextOutputMultiplier:    openAIGPT54LongContextOutputMultiplier,
 	}
 	// OpenAI GPT-5.5 官方价格；Fast 为标准价 2.5 倍。
 	// Source: https://platform.openai.com/docs/pricing
@@ -405,24 +396,18 @@ func (s *BillingService) initFallbackPricing() {
 		InputPricePerToken:  5e-6,
 		OutputPricePerToken: 30e-6,
 		// 官方未列独立 cache-write 价；内部出现 cache creation token 时按输入价兜底。
-		CacheCreationPricePerToken:  5e-6,
-		CacheReadPricePerToken:      0.5e-6,
-		SupportsCacheBreakdown:      false,
-		LongContextInputThreshold:   openAIGPT54LongContextInputThreshold,
-		LongContextInputMultiplier:  openAIGPT54LongContextInputMultiplier,
-		LongContextOutputMultiplier: openAIGPT54LongContextOutputMultiplier,
+		CacheCreationPricePerToken: 5e-6,
+		CacheReadPricePerToken:     0.5e-6,
+		SupportsCacheBreakdown:     false,
 	}, 2.5)
 	// GPT-5.5 Pro 当前不提供 Fast；保留标准、Flex 和长上下文 fallback 价格。
 	s.fallbackPrices["gpt-5.5-pro"] = &ModelPricing{
 		InputPricePerToken:  30e-6,
 		OutputPricePerToken: 180e-6,
 		// 官方未列独立 cached-input/cache-write 价；内部出现对应 token 时按输入价兜底。
-		CacheCreationPricePerToken:  30e-6,
-		CacheReadPricePerToken:      30e-6,
-		SupportsCacheBreakdown:      false,
-		LongContextInputThreshold:   openAIGPT54LongContextInputThreshold,
-		LongContextInputMultiplier:  openAIGPT54LongContextInputMultiplier,
-		LongContextOutputMultiplier: openAIGPT54LongContextOutputMultiplier,
+		CacheCreationPricePerToken: 30e-6,
+		CacheReadPricePerToken:     30e-6,
+		SupportsCacheBreakdown:     false,
 	}

 	// OpenAI GPT-5.6 官方价格（USD/token）。缓存写入为输入价的 1.25 倍。
@@ -435,9 +420,6 @@ func (s *BillingService) initFallbackPricing() {
 		CacheCreationPricePerTokenPriority: 12.5e-6,
 		CacheReadPricePerToken:             0.5e-6,
 		CacheReadPricePerTokenPriority:     1e-6,
-		LongContextInputThreshold:          openAIGPT54LongContextInputThreshold,
-		LongContextInputMultiplier:         openAIGPT54LongContextInputMultiplier,
-		LongContextOutputMultiplier:        openAIGPT54LongContextOutputMultiplier,
 	}
 	s.fallbackPrices["gpt-5.6-terra"] = &ModelPricing{
 		InputPricePerToken:                 2e-6,
@@ -448,9 +430,6 @@ func (s *BillingService) initFallbackPricing() {
 		CacheCreationPricePerTokenPriority: 5e-6,
 		CacheReadPricePerToken:             0.2e-6,
 		CacheReadPricePerTokenPriority:     0.4e-6,
-		LongContextInputThreshold:          openAIGPT54LongContextInputThreshold,
-		LongContextInputMultiplier:         openAIGPT54LongContextInputMultiplier,
-		LongContextOutputMultiplier:        openAIGPT54LongContextOutputMultiplier,
 	}
 	s.fallbackPrices["gpt-5.6-luna"] = &ModelPricing{
 		InputPricePerToken:                 0.2e-6,
@@ -461,9 +440,6 @@ func (s *BillingService) initFallbackPricing() {
 		CacheCreationPricePerTokenPriority: 0.5e-6,
 		CacheReadPricePerToken:             0.02e-6,
 		CacheReadPricePerTokenPriority:     0.04e-6,
-		LongContextInputThreshold:          openAIGPT54LongContextInputThreshold,
-		LongContextInputMultiplier:         openAIGPT54LongContextInputMultiplier,
-		LongContextOutputMultiplier:        openAIGPT54LongContextOutputMultiplier,
 	}

 	s.fallbackPrices["gpt-5.4-mini"] = &ModelPricing{
@@ -1125,10 +1101,12 @@ func (s *BillingService) GetModelPricing(model string) (*ModelPricing, error) {
 				CacheCreation1hPrice:               price1h,
 				SupportsCacheBreakdown:             enableBreakdown,
 				LongContextInputThreshold:          litellmPricing.LongContextInputTokenThreshold,
-				LongContextInputMultiplier:         litellmPricing.LongContextInputCostMultiplier,
-				LongContextOutputMultiplier:        litellmPricing.LongContextOutputCostMultiplier,
-				ImageInputPricePerToken:            litellmPricing.InputCostPerImageToken,
-				ImageOutputPricePerToken:           litellmPricing.OutputCostPerImageToken,
+				// xAI 的长上下文阈值语义为"达到即进高档"（LiteLLM 同口径），其余提供商为严格大于。
+				LongContextThresholdInclusive: strings.EqualFold(litellmPricing.LiteLLMProvider, "xai"),
+				LongContextInputMultiplier:    litellmPricing.LongContextInputCostMultiplier,
+				LongContextOutputMultiplier:   litellmPricing.LongContextOutputCostMultiplier,
+				ImageInputPricePerToken:       litellmPricing.InputCostPerImageToken,
+				ImageOutputPricePerToken:      litellmPricing.OutputCostPerImageToken,
 			}), nil
 		}
 	}
@@ -1375,15 +1353,18 @@ func (s *BillingService) computeTokenBreakdown(
 	var baselineCost *CostBreakdown
 	if longContextPricingEligible {
 		baselineCost = s.computeTokenBreakdown(pricing, tokens, rateMultiplier, serviceTier, false)
-		inputPrice *= pricing.LongContextInputMultiplier
-		outputPrice *= pricing.LongContextOutputMultiplier
+		// 倍率 ≤0 表示该项未配置（目录/覆写条目可能只写了 input 或 output 一侧），
+		// 按 1 计而不是乘 0：乘 0 会把超阈值请求的对应分项算成免费。
+		longCtxInputMultiplier := longContextMultiplierOrOne(pricing.LongContextInputMultiplier)
+		inputPrice *= longCtxInputMultiplier
+		outputPrice *= longContextMultiplierOrOne(pricing.LongContextOutputMultiplier)
 		// 缓存读取本质上是输入侧的复用，应与 input 一同应用长上下文倍率；
 		// 否则 cache hit 越多，少计的费用越多（见 #2293）。
-		cacheReadPrice *= pricing.LongContextInputMultiplier
+		cacheReadPrice *= longCtxInputMultiplier
 		// 缓存创建（cache_write）也是输入侧操作，三档价格（标准 / 5m / 1h）
 		// 都通过 computeCacheCreationCost 直接读取 pricing.*，不会经过这里
 		// 的倍率修改，因此显式向下传一个倍率，避免长上下文场景下被漏乘。
-		cacheCreationMultiplier = pricing.LongContextInputMultiplier
+		cacheCreationMultiplier = longCtxInputMultiplier
 	}

 	bd := &CostBreakdown{}
@@ -1568,10 +1549,13 @@ func (s *BillingService) calculateCostInternalWithPolicy(
 	return s.computeTokenBreakdown(pricing, tokens, rateMultiplier, serviceTier, longContextBillingEnabled), nil
 }

-// applyModelSpecificPricingPolicy 应用模型特定定价策略（GPT-5.6 长上下文/缓存写入、
-// DeepSeek 官方价强制覆盖等）。默认强制 DeepSeek 官方价——该路径仅被默认价卡
-// （GetModelPricing 内部）调用；分组/渠道自定义定价路径用带参数的
-// applyModelSpecificPricingPolicyEx 关闭强制，保留运营者配置。
+// applyModelSpecificPricingPolicy 对目录数据做模型特定修正：DeepSeek 官方价
+// 强制覆盖；GPT-5.6 缺 cache_write 价时按官方规则补 1.25 倍输入价；Fast/priority
+// 档按业务倍率改写（本地/远程目录的 priority 价可能沿用官方旧口径）。长上下文
+// 阶梯不在此处补齐：一律由目录数据（above_XXXk 折算或显式 long_context_* 字段）
+// 驱动。默认强制 DeepSeek 官方价——该路径仅被默认价卡（GetModelPricing 内部）
+// 调用；分组/渠道自定义定价路径用带参数的 applyModelSpecificPricingPolicyEx
+// 关闭强制，保留运营者配置。
 func (s *BillingService) applyModelSpecificPricingPolicy(model string, pricing *ModelPricing) *ModelPricing {
 	return s.applyModelSpecificPricingPolicyEx(model, pricing, true)
 }
@@ -1607,16 +1591,10 @@ func (s *BillingService) applyModelSpecificPricingPolicyEx(model string, pricing
 	}
 	normalized := normalizeKnownOpenAICodexModel(model)
 	isGPT56 := isOpenAIGPT56Model(normalized)
-	usesLegacyLongContextPricing := usesOpenAILegacyLongContextPricing(normalized)
-	if !isGPT56 && !usesLegacyLongContextPricing {
-		return pricing
-	}
-	needsLongContextPolicy := (isGPT56 || usesLegacyLongContextPricing) &&
-		(pricing.LongContextInputThreshold <= 0 || pricing.LongContextInputMultiplier <= 0 || pricing.LongContextOutputMultiplier <= 0)
 	needsCacheCreationPolicy := isGPT56 && !pricing.CacheCreationPriceExplicit && (pricing.CacheCreationPricePerToken <= 0 ||
 		(pricing.InputPricePerTokenPriority > 0 && pricing.CacheCreationPricePerTokenPriority <= 0))
 	fastRatio := openAIModelFastPricingRatio(normalized)
-	if !needsLongContextPolicy && !needsCacheCreationPolicy && fastRatio <= 0 {
+	if !needsCacheCreationPolicy && fastRatio <= 0 {
 		return pricing
 	}
 	cloned := *pricing
@@ -1628,17 +1606,6 @@ func (s *BillingService) applyModelSpecificPricingPolicyEx(model string, pricing
 			cloned.CacheCreationPricePerTokenPriority = cloned.InputPricePerTokenPriority * 1.25
 		}
 	}
-	if isGPT56 || usesLegacyLongContextPricing {
-		if cloned.LongContextInputThreshold <= 0 {
-			cloned.LongContextInputThreshold = openAIGPT54LongContextInputThreshold
-		}
-		if cloned.LongContextInputMultiplier <= 0 {
-			cloned.LongContextInputMultiplier = openAIGPT54LongContextInputMultiplier
-		}
-		if cloned.LongContextOutputMultiplier <= 0 {
-			cloned.LongContextOutputMultiplier = openAIGPT54LongContextOutputMultiplier
-		}
-	}
 	if fastRatio > 0 {
 		enforceOpenAIFastPricingRatio(&cloned, fastRatio)
 	}
@@ -1678,6 +1645,14 @@ func enforceOpenAIFastPricingRatio(pricing *ModelPricing, ratio float64) {
 	}
 }

+// longContextMultiplierOrOne 把未配置（≤0）的长上下文倍率归一为 1。
+func longContextMultiplierOrOne(m float64) float64 {
+	if m <= 0 {
+		return 1
+	}
+	return m
+}
+
 func (s *BillingService) shouldApplySessionLongContextPricing(tokens UsageTokens, pricing *ModelPricing) bool {
 	if pricing == nil || pricing.LongContextInputThreshold <= 0 {
 		return false
@@ -1692,10 +1667,6 @@ func (s *BillingService) shouldApplySessionLongContextPricing(tokens UsageTokens
 	return totalInputTokens > pricing.LongContextInputThreshold
 }

-func usesOpenAILegacyLongContextPricing(normalized string) bool {
-	return normalized == "gpt-5.4" || normalized == "gpt-5.5" || normalized == "gpt-5.5-pro"
-}
-
 // CalculateCostWithConfig 使用配置中的默认倍率计算费用
 func (s *BillingService) CalculateCostWithConfig(model string, tokens UsageTokens) (*CostBreakdown, error) {
 	multiplier := s.cfg.Default.RateMultiplier
@@ -1705,82 +1676,6 @@ func (s *BillingService) CalculateCostWithConfig(model string, tokens UsageToken
 	return s.CalculateCost(model, tokens, multiplier)
 }

-// CalculateCostWithLongContext 计算费用，支持长上下文双倍计费
-// threshold: 阈值（如 200000），超过此值的部分按 extraMultiplier 倍计费
-// extraMultiplier: 超出部分的倍率（如 2.0 表示双倍）
-//
-// 示例：缓存 210k + 输入 10k = 220k，阈值 200k，倍率 2.0
-// 拆分为：范围内 (200k, 0) + 范围外 (10k, 10k)
-// 范围内正常计费，范围外 × 2 计费
-func (s *BillingService) CalculateCostWithLongContext(model string, tokens UsageTokens, rateMultiplier float64, threshold int, extraMultiplier float64) (*CostBreakdown, error) {
-	// 未启用长上下文计费，直接走正常计费
-	if threshold <= 0 || extraMultiplier <= 1 {
-		return s.CalculateCost(model, tokens, rateMultiplier)
-	}
-
-	// 计算总输入 token（缓存读取 + 新输入）
-	total := tokens.CacheReadTokens + tokens.InputTokens
-	if total <= threshold {
-		return s.CalculateCost(model, tokens, rateMultiplier)
-	}
-
-	// 拆分成范围内和范围外
-	var inRangeCacheTokens, inRangeInputTokens int
-	var outRangeCacheTokens, outRangeInputTokens int
-
-	if tokens.CacheReadTokens >= threshold {
-		// 缓存已超过阈值：范围内只有缓存，范围外是超出的缓存+全部输入
-		inRangeCacheTokens = threshold
-		inRangeInputTokens = 0
-		outRangeCacheTokens = tokens.CacheReadTokens - threshold
-		outRangeInputTokens = tokens.InputTokens
-	} else {
-		// 缓存未超过阈值：范围内是全部缓存+部分输入，范围外是剩余输入
-		inRangeCacheTokens = tokens.CacheReadTokens
-		inRangeInputTokens = threshold - tokens.CacheReadTokens
-		outRangeCacheTokens = 0
-		outRangeInputTokens = tokens.InputTokens - inRangeInputTokens
-	}
-
-	// 范围内部分：正常计费
-	inRangeTokens := UsageTokens{
-		InputTokens:           inRangeInputTokens,
-		OutputTokens:          tokens.OutputTokens, // 输出只算一次
-		CacheCreationTokens:   tokens.CacheCreationTokens,
-		CacheReadTokens:       inRangeCacheTokens,
-		CacheCreation5mTokens: tokens.CacheCreation5mTokens,
-		CacheCreation1hTokens: tokens.CacheCreation1hTokens,
-		ImageOutputTokens:     tokens.ImageOutputTokens,
-	}
-	inRangeCost, err := s.CalculateCost(model, inRangeTokens, rateMultiplier)
-	if err != nil {
-		return nil, err
-	}
-
-	// 范围外部分：× extraMultiplier 计费
-	outRangeTokens := UsageTokens{
-		InputTokens:     outRangeInputTokens,
-		CacheReadTokens: outRangeCacheTokens,
-	}
-	outRangeCost, err := s.CalculateCost(model, outRangeTokens, rateMultiplier*extraMultiplier)
-	if err != nil {
-		return inRangeCost, fmt.Errorf("out-range cost: %w", err)
-	}
-
-	// 合并成本
-	return &CostBreakdown{
-		InputCost:                 inRangeCost.InputCost + outRangeCost.InputCost,
-		ImageInputCost:            inRangeCost.ImageInputCost + outRangeCost.ImageInputCost,
-		OutputCost:                inRangeCost.OutputCost,
-		ImageOutputCost:           inRangeCost.ImageOutputCost,
-		CacheCreationCost:         inRangeCost.CacheCreationCost,
-		CacheReadCost:             inRangeCost.CacheReadCost + outRangeCost.CacheReadCost,
-		TotalCost:                 inRangeCost.TotalCost + outRangeCost.TotalCost,
-		ActualCost:                inRangeCost.ActualCost + outRangeCost.ActualCost,
-		LongContextBillingApplied: outRangeCost.ActualCost > 0,
-	}, nil
-}
-
 // ListSupportedModels 列出所有支持的模型（现在总是返回true，因为有模糊匹配）
 func (s *BillingService) ListSupportedModels() []string {
 	models := make([]string, 0)
@@ -33,6 +33,11 @@ func newTestBillingService() *BillingService {
 	return NewBillingService(&config.Config{}, nil)
 }

+func newTestBillingServiceWithOpenAILadderCatalog(t *testing.T) *BillingService {
+	t.Helper()
+	return NewBillingService(&config.Config{}, newStubPricingServiceFromJSON(t, openAILadderCatalogJSON))
+}
+
 func TestCalculateCost_BasicComputation(t *testing.T) {
 	svc := newTestBillingService()

@@ -200,9 +205,21 @@ func TestGetModelPricing_OpenAIGPT54Fallback(t *testing.T) {
 	require.InDelta(t, 2.5e-6, pricing.InputPricePerToken, 1e-12)
 	require.InDelta(t, 15e-6, pricing.OutputPricePerToken, 1e-12)
 	require.InDelta(t, 0.25e-6, pricing.CacheReadPricePerToken, 1e-12)
+	// 静态兜底价不携带长上下文阶梯：阶梯一律由目录数据（above_272k 折算）驱动。
+	require.Zero(t, pricing.LongContextInputThreshold)
+	require.Zero(t, pricing.LongContextInputMultiplier)
+	require.Zero(t, pricing.LongContextOutputMultiplier)
+}
+
+func TestGetModelPricing_CatalogAboveTierFieldsDriveLongContext(t *testing.T) {
+	svc := newTestBillingServiceWithOpenAILadderCatalog(t)
+
+	pricing, err := svc.GetModelPricing("gpt-5.4")
+	require.NoError(t, err)
 	require.Equal(t, 272000, pricing.LongContextInputThreshold)
 	require.InDelta(t, 2.0, pricing.LongContextInputMultiplier, 1e-12)
 	require.InDelta(t, 1.5, pricing.LongContextOutputMultiplier, 1e-12)
+	require.False(t, pricing.LongContextThresholdInclusive, "openai 阈值语义为严格大于")
 }

 func TestGetModelPricing_OpenAICompactAliasesFallback(t *testing.T) {
@@ -215,8 +232,8 @@ func TestGetModelPricing_OpenAICompactAliasesFallback(t *testing.T) {
 		cacheRead   float64
 		longContext int
 	}{
-		{model: "gpt5.5", inputPrice: 5e-6, outputPrice: 30e-6, cacheRead: 0.5e-6, longContext: 272000},
-		{model: "openai/gpt5.4", inputPrice: 2.5e-6, outputPrice: 15e-6, cacheRead: 0.25e-6, longContext: 272000},
+		{model: "gpt5.5", inputPrice: 5e-6, outputPrice: 30e-6, cacheRead: 0.5e-6, longContext: 0},
+		{model: "openai/gpt5.4", inputPrice: 2.5e-6, outputPrice: 15e-6, cacheRead: 0.25e-6, longContext: 0},
 		{model: "gpt5.4-mini", inputPrice: 7.5e-7, outputPrice: 4.5e-6, cacheRead: 7.5e-8, longContext: 0},
 		{model: "gpt5.3codexspark", inputPrice: 1.5e-6, outputPrice: 12e-6, cacheRead: 0.15e-6, longContext: 0},
 	}
@@ -247,7 +264,7 @@ func TestGetModelPricing_OpenAIGPT54MiniFallback(t *testing.T) {
 }

 func TestCalculateCost_OpenAIGPT54LongContextAppliesWholeSessionMultipliers(t *testing.T) {
-	svc := newTestBillingService()
+	svc := newTestBillingServiceWithOpenAILadderCatalog(t)

 	tokens := UsageTokens{
 		InputTokens:  300000,
@@ -267,7 +284,7 @@ func TestCalculateCost_OpenAIGPT54LongContextAppliesWholeSessionMultipliers(t *t
 }

 func TestCalculateCost_OpenAIGPT54LongContextMarkerRequiresActualCostIncrease(t *testing.T) {
-	svc := newTestBillingService()
+	svc := newTestBillingServiceWithOpenAILadderCatalog(t)

 	cost, err := svc.calculateCostWithServiceTierPolicy(
 		"gpt-5.4-2026-03-05",
@@ -283,7 +300,7 @@ func TestCalculateCost_OpenAIGPT54LongContextMarkerRequiresActualCostIncrease(t
 }

 func TestCalculateCost_OpenAIGPT55ProUsesGPT55PricingPolicy(t *testing.T) {
-	svc := newTestBillingService()
+	svc := newTestBillingServiceWithOpenAILadderCatalog(t)

 	tokens := UsageTokens{
 		InputTokens:  300000,
@@ -331,7 +348,7 @@ func TestFallbackPricing_OpenAIGPT55ProUsesOfficialPrices(t *testing.T) {
 // 修复前：CacheReadCost = tokens * 0.25e-6 （漏乘倍率，少计费用）。
 // 修复后：CacheReadCost = tokens * 0.25e-6 * LongContextInputMultiplier(=2.0)。
 func TestCalculateCost_OpenAIGPT54LongContextAppliesMultiplierToCacheRead(t *testing.T) {
-	svc := newTestBillingService()
+	svc := newTestBillingServiceWithOpenAILadderCatalog(t)

 	// InputTokens + CacheReadTokens = 1000 + 300000 = 301000 > 272000 阈值
 	tokens := UsageTokens{
@@ -359,7 +376,7 @@ func TestCalculateCost_OpenAIGPT54LongContextAppliesMultiplierToCacheRead(t *tes

 // 阴性测试：未触发长上下文时，cache_read_price 不应被错误地乘以倍率。
 func TestCalculateCost_OpenAIGPT54NoLongContextKeepsCacheReadAtBasePrice(t *testing.T) {
-	svc := newTestBillingService()
+	svc := newTestBillingServiceWithOpenAILadderCatalog(t)

 	// InputTokens + CacheReadTokens = 1000 + 100000 = 101000 < 272000 阈值，不触发长上下文
 	tokens := UsageTokens{
@@ -381,7 +398,7 @@ func TestCalculateCost_OpenAIGPT54NoLongContextKeepsCacheReadAtBasePrice(t *test
 // 不经过 computeTokenBreakdown 内的 inputPrice / cacheReadPrice 倍率修改，因此
 // 修复前 cache_creation 部分会按基础价计算，少计费用约 50%（默认倍率 2.0）。
 func TestCalculateCost_OpenAIGPT54LongContextAppliesMultiplierToCacheCreation(t *testing.T) {
-	svc := newTestBillingService()
+	svc := newTestBillingServiceWithOpenAILadderCatalog(t)

 	// InputTokens + CacheReadTokens = 1000 + 300000 = 301000 > 272000 阈值
 	tokens := UsageTokens{
@@ -402,7 +419,7 @@ func TestCalculateCost_OpenAIGPT54LongContextAppliesMultiplierToCacheCreation(t

 // 阴性测试：未触发长上下文时，cache_creation_price 不应被错误地乘以倍率。
 func TestCalculateCost_OpenAIGPT54NoLongContextKeepsCacheCreationAtBasePrice(t *testing.T) {
-	svc := newTestBillingService()
+	svc := newTestBillingServiceWithOpenAILadderCatalog(t)

 	// InputTokens + CacheReadTokens = 1000 + 100000 = 101000 < 272000 阈值，不触发长上下文
 	tokens := UsageTokens{
@@ -921,114 +938,6 @@ func TestComputeTokenBreakdown_GptImage2ImageEditIssue4386(t *testing.T) {
 	require.InDelta(t, wantImageOutput, cost.ImageOutputCost, 1e-15)
 	require.InDelta(t, 0.016081, cost.TotalCost, 1e-9, "总额应为 $0.016081（修复前为 $0.015025）")
 }
-func TestCalculateCostWithLongContext_BelowThreshold(t *testing.T) {
-	svc := newTestBillingService()
-
-	tokens := UsageTokens{
-		InputTokens:     50000,
-		OutputTokens:    1000,
-		CacheReadTokens: 100000,
-	}
-	// 总输入 150k < 200k 阈值，应走正常计费
-	cost, err := svc.CalculateCostWithLongContext("claude-sonnet-4", tokens, 1.0, 200000, 2.0)
-	require.NoError(t, err)
-
-	normalCost, err := svc.CalculateCost("claude-sonnet-4", tokens, 1.0)
-	require.NoError(t, err)
-
-	require.InDelta(t, normalCost.ActualCost, cost.ActualCost, 1e-10)
-}
-
-func TestCalculateCostWithLongContext_AboveThreshold_CacheExceedsThreshold(t *testing.T) {
-	svc := newTestBillingService()
-
-	// 缓存 210k + 输入 10k = 220k > 200k 阈值
-	// 缓存已超阈值：范围内 200k 缓存，范围外 10k 缓存 + 10k 输入
-	tokens := UsageTokens{
-		InputTokens:     10000,
-		OutputTokens:    1000,
-		CacheReadTokens: 210000,
-	}
-	cost, err := svc.CalculateCostWithLongContext("claude-sonnet-4", tokens, 1.0, 200000, 2.0)
-	require.NoError(t, err)
-
-	// 范围内：200k cache + 0 input + 1k output
-	inRange, _ := svc.CalculateCost("claude-sonnet-4", UsageTokens{
-		InputTokens:     0,
-		OutputTokens:    1000,
-		CacheReadTokens: 200000,
-	}, 1.0)
-
-	// 范围外：10k cache + 10k input，倍率 2.0
-	outRange, _ := svc.CalculateCost("claude-sonnet-4", UsageTokens{
-		InputTokens:     10000,
-		CacheReadTokens: 10000,
-	}, 2.0)
-
-	require.InDelta(t, inRange.ActualCost+outRange.ActualCost, cost.ActualCost, 1e-10)
-}
-
-func TestCalculateCostWithLongContext_AboveThreshold_CacheBelowThreshold(t *testing.T) {
-	svc := newTestBillingService()
-
-	// 缓存 100k + 输入 150k = 250k > 200k 阈值
-	// 缓存未超阈值：范围内 100k 缓存 + 100k 输入，范围外 50k 输入
-	tokens := UsageTokens{
-		InputTokens:     150000,
-		OutputTokens:    1000,
-		CacheReadTokens: 100000,
-	}
-	cost, err := svc.CalculateCostWithLongContext("claude-sonnet-4", tokens, 1.0, 200000, 2.0)
-	require.NoError(t, err)
-
-	require.True(t, cost.ActualCost > 0, "费用应大于 0")
-
-	// 正常费用不含长上下文
-	normalCost, _ := svc.CalculateCost("claude-sonnet-4", tokens, 1.0)
-	require.True(t, cost.ActualCost > normalCost.ActualCost, "长上下文费用应高于正常费用")
-}
-
-func TestCalculateCostWithLongContext_MarkerRequiresActualCostIncrease(t *testing.T) {
-	svc := newTestBillingService()
-	tokens := UsageTokens{InputTokens: 300000}
-
-	cost, err := svc.CalculateCostWithLongContext("claude-sonnet-4", tokens, 0, 200000, 2.0)
-
-	require.NoError(t, err)
-	require.Zero(t, cost.ActualCost)
-	require.False(t, cost.LongContextBillingApplied)
-}
-
-func TestCalculateCostWithLongContext_DisabledThreshold(t *testing.T) {
-	svc := newTestBillingService()
-
-	tokens := UsageTokens{InputTokens: 300000, CacheReadTokens: 0}
-
-	// threshold <= 0 应禁用长上下文计费
-	cost1, err := svc.CalculateCostWithLongContext("claude-sonnet-4", tokens, 1.0, 0, 2.0)
-	require.NoError(t, err)
-
-	cost2, err := svc.CalculateCost("claude-sonnet-4", tokens, 1.0)
-	require.NoError(t, err)
-
-	require.InDelta(t, cost2.ActualCost, cost1.ActualCost, 1e-10)
-}
-
-func TestCalculateCostWithLongContext_ExtraMultiplierLessEqualOne(t *testing.T) {
-	svc := newTestBillingService()
-
-	tokens := UsageTokens{InputTokens: 300000}
-
-	// extraMultiplier <= 1 应禁用长上下文计费
-	cost, err := svc.CalculateCostWithLongContext("claude-sonnet-4", tokens, 1.0, 200000, 1.0)
-	require.NoError(t, err)
-
-	normalCost, err := svc.CalculateCost("claude-sonnet-4", tokens, 1.0)
-	require.NoError(t, err)
-
-	require.InDelta(t, normalCost.ActualCost, cost.ActualCost, 1e-10)
-}
-
 func TestCalculateImageCost(t *testing.T) {
 	svc := newTestBillingService()

@@ -1178,19 +1087,6 @@ func TestForceUpdatePricing_NilService(t *testing.T) {
 	require.Contains(t, err.Error(), "not initialized")
 }

-func TestCalculateCostWithLongContext_PropagatesError(t *testing.T) {
-	// 使用空的 fallback prices 让 GetModelPricing 失败
-	svc := &BillingService{
-		cfg:            &config.Config{},
-		fallbackPrices: make(map[string]*ModelPricing),
-	}
-
-	tokens := UsageTokens{InputTokens: 300000, CacheReadTokens: 0}
-	_, err := svc.CalculateCostWithLongContext("unknown-model", tokens, 1.0, 200000, 2.0)
-	require.Error(t, err)
-	require.Contains(t, err.Error(), "pricing not found")
-}
-
 func TestGetModelPricing_Grok45OfficialFallback(t *testing.T) {
 	svc := newTestBillingService()

@@ -5,33 +5,6 @@ import (
 	"time"
 )

-// LegacyLongContextRule 平台级"超出阈值部分按倍率计费"的旧规则。
-//
-// 语义为边际计费：仅 input/cache_read 中超过 Threshold 的部分乘 Multiplier，
-// output 与 cache_write 不受影响（见 CalculateCostWithLongContext）。
-// 只有 Gemini 原生 /v1beta 入口在该分组对该模型没有分组/渠道定价时才使用；
-// 规则常量由 BillingService 统一持有，网关与模型广场都从这里读取。
-type LegacyLongContextRule struct {
-	Threshold  int
-	Multiplier float64
-}
-
-const (
-	geminiLegacyLongContextThreshold  = 200000
-	geminiLegacyLongContextMultiplier = 2.0
-)
-
-// LegacyLongContextRule 返回平台的旧长上下文规则；无规则的平台返回 nil。
-func (s *BillingService) LegacyLongContextRule(platform string) *LegacyLongContextRule {
-	if platform == PlatformGemini {
-		return &LegacyLongContextRule{
-			Threshold:  geminiLegacyLongContextThreshold,
-			Multiplier: geminiLegacyLongContextMultiplier,
-		}
-	}
-	return nil
-}
-
 // TokenCostRequest 通用网关 token 计费请求。
 type TokenCostRequest struct {
 	Ctx            context.Context
@@ -44,27 +17,12 @@ type TokenCostRequest struct {
 	Resolver       *ModelPricingResolver
 	// Resolved 为调用方预先解析的定价（Resolver.Resolve 的结果），nil 表示未解析。
 	Resolved *ResolvedPricing
-	// LegacyLongContext 入口携带的旧长上下文规则，nil 表示该入口不使用。
-	LegacyLongContext *LegacyLongContextRule
-}
-
-// legacyLongContextApplies 判定请求是否走旧长上下文规则：
-// 分组/渠道显式定价优先；否则在规则存在且分组长上下文开关开启时生效。
-func legacyLongContextApplies(resolved *ResolvedPricing, group *Group, rule *LegacyLongContextRule) bool {
-	if rule == nil || rule.Threshold <= 0 {
-		return false
-	}
-	if resolved != nil && (resolved.Source == PricingSourceGroup || resolved.Source == PricingSourceChannel) {
-		return false
-	}
-	return group == nil || group.LongContextPricingEnabled
 }

 // CalculateTokenCostForRequest 按通用网关的路径选择计算 token 费用：
-//  1. 分组/渠道显式定价 → 统一计费（区间、分组卡、目录阶梯均在其中）；
-//  2. 否则入口带旧长上下文规则且分组开关开启 → 旧边际计费；
-//  3. 否则有解析器与分组 → 统一计费（内置目录定价）；
-//  4. 否则按模型目录直接计费。
+//  1. 分组/渠道显式定价，或有解析器与分组 → 统一计费
+//     （区间、分组卡、目录长上下文阶梯均在其中，阶梯由目录数据驱动）；
+//  2. 否则按模型目录直接计费。
 //
 // 模型广场的阶梯表查询与网关使用同一入口，保证展示与扣费同源。
 func (s *BillingService) CalculateTokenCostForRequest(req TokenCostRequest) (*CostBreakdown, error) {
@@ -72,10 +30,6 @@ func (s *BillingService) CalculateTokenCostForRequest(req TokenCostRequest) (*Co
 	if resolved != nil && (resolved.Source == PricingSourceGroup || resolved.Source == PricingSourceChannel) {
 		return s.CalculateCostUnified(s.tokenCostInput(req, resolved))
 	}
-	if legacyLongContextApplies(resolved, req.Group, req.LegacyLongContext) {
-		return s.CalculateCostWithLongContext(req.Model, req.Tokens, req.RateMultiplier,
-			req.LegacyLongContext.Threshold, req.LegacyLongContext.Multiplier)
-	}
 	if req.Resolver != nil && req.Group != nil {
 		return s.CalculateCostUnified(s.tokenCostInput(req, resolved))
 	}
@@ -28,6 +28,7 @@ func newTokenCostTestEnv(t *testing.T, groupPlatform string, pricing []ChannelMo
 	return bs, NewModelPricingResolver(cs, bs)
 }

+// geminiCatalogStub 无阶梯字段的 gemini 目录条目（用于验证"无数据即无阶梯"）。
 func geminiCatalogStub() *PricingService {
 	return newStubPricingServiceFromMap(map[string]*LiteLLMModelPricing{
 		"gemini-2.5-pro": {
@@ -39,22 +40,29 @@ func geminiCatalogStub() *PricingService {
 	})
 }

-func TestLegacyLongContextRule_OnlyGemini(t *testing.T) {
-	bs := NewBillingService(&config.Config{}, nil)
-	rule := bs.LegacyLongContextRule(PlatformGemini)
-	require.NotNil(t, rule)
-	require.Equal(t, 200000, rule.Threshold)
-	require.InDelta(t, 2.0, rule.Multiplier, 1e-12)
-	for _, platform := range []string{PlatformOpenAI, PlatformAnthropic, PlatformAntigravity, PlatformComposite, ""} {
-		require.Nil(t, bs.LegacyLongContextRule(platform), platform)
-	}
+// geminiLadderCatalogJSON 镜像真实目录的 gemini pro 条目：above_200k 绝对价，
+// 由解析层折算成 200K 阈值 + 输入 ×2 / 输出 ×1.5。
+const geminiLadderCatalogJSON = `{
+	"gemini-2.5-pro": {"litellm_provider": "vertex_ai-language-models", "mode": "chat",
+		"input_cost_per_token": 1.25e-06, "output_cost_per_token": 1e-05,
+		"cache_read_input_token_cost": 3.125e-07,
+		"input_cost_per_token_above_200k_tokens": 2.5e-06,
+		"output_cost_per_token_above_200k_tokens": 1.5e-05,
+		"cache_read_input_token_cost_above_200k_tokens": 6.25e-07}
+}`
+
+func geminiLadderCatalogStub(t *testing.T) *PricingService {
+	t.Helper()
+	return newStubPricingServiceFromJSON(t, geminiLadderCatalogJSON)
 }

-func TestCalculateTokenCostForRequest_ChannelPricingWinsOverLegacyRule(t *testing.T) {
+// 渠道平价之上叠加目录阶梯：与分组价卡/OpenAI 渠道价的既有语义一致，
+// 超阈值整单按渠道价 × 目录倍率。
+func TestCalculateTokenCostForRequest_ChannelFlatPriceStacksCatalogLadder(t *testing.T) {
 	bs, resolver := newTokenCostTestEnv(t, PlatformGemini, []ChannelModelPricing{{
 		Platform: PlatformGemini, Models: []string{"gemini-2.5-pro"}, BillingMode: BillingModeToken,
 		InputPrice: testPtrFloat64(10e-6), OutputPrice: testPtrFloat64(40e-6),
-	}}, geminiCatalogStub())
+	}}, geminiLadderCatalogStub(t))
 	group := &Group{ID: 100, Platform: PlatformGemini, LongContextPricingEnabled: true}
 	gid := group.ID
 	resolved := resolver.Resolve(context.Background(), PricingInput{Model: "gemini-2.5-pro", GroupID: &gid, Group: group})
@@ -63,23 +71,41 @@ func TestCalculateTokenCostForRequest_ChannelPricingWinsOverLegacyRule(t *testin
 	tokens := UsageTokens{InputTokens: 300000, OutputTokens: 1000}
 	got, err := bs.CalculateTokenCostForRequest(TokenCostRequest{
 		Ctx: context.Background(), Model: "gemini-2.5-pro", Group: group, Tokens: tokens, RateMultiplier: 1,
-		Resolver: resolver, Resolved: resolved, LegacyLongContext: bs.LegacyLongContextRule(PlatformGemini),
+		Resolver: resolver, Resolved: resolved,
 	})
 	require.NoError(t, err)
-	want, err := bs.CalculateCostUnified(CostInput{
-		Ctx: context.Background(), Model: "gemini-2.5-pro", GroupID: &gid, Group: group, Tokens: tokens,
-		RequestCount: 1, RateMultiplier: 1, Resolver: resolver, Resolved: resolved,
-	})
-	require.NoError(t, err)
-	require.Equal(t, want, got)
-	// 渠道平价 10e-6 × 300K，旧规则未叠加
-	require.InDelta(t, 3.0, got.InputCost, 1e-9)
+	require.InDelta(t, 300000*10e-6*2, got.InputCost, 1e-9)
+	require.InDelta(t, 1000*40e-6*1.5, got.OutputCost, 1e-9)
+	require.True(t, got.LongContextBillingApplied)
 }

-func TestCalculateTokenCostForRequest_LegacyRuleFollowsGroupToggle(t *testing.T) {
-	bs, resolver := newTokenCostTestEnv(t, PlatformGemini, nil, geminiCatalogStub())
+// 渠道配置了定价区间时以渠道区间为准：目录阶梯（倍率）不再叠加。
+func TestCalculateTokenCostForRequest_ChannelIntervalsOverrideCatalogLadder(t *testing.T) {
+	bs, resolver := newTokenCostTestEnv(t, PlatformGemini, []ChannelModelPricing{{
+		Platform: PlatformGemini, Models: []string{"gemini-2.5-pro"}, BillingMode: BillingModeToken,
+		Intervals: []PricingInterval{{MinTokens: 0, InputPrice: testPtrFloat64(10e-6), OutputPrice: testPtrFloat64(40e-6)}},
+	}}, geminiLadderCatalogStub(t))
+	group := &Group{ID: 100, Platform: PlatformGemini, LongContextPricingEnabled: true}
+	gid := group.ID
+	resolved := resolver.Resolve(context.Background(), PricingInput{Model: "gemini-2.5-pro", GroupID: &gid, Group: group})
+	require.Equal(t, PricingSourceChannel, resolved.Source)
+	require.NotEmpty(t, resolved.Intervals)
+
+	tokens := UsageTokens{InputTokens: 300000, OutputTokens: 1000}
+	got, err := bs.CalculateTokenCostForRequest(TokenCostRequest{
+		Ctx: context.Background(), Model: "gemini-2.5-pro", Group: group, Tokens: tokens, RateMultiplier: 1,
+		Resolver: resolver, Resolved: resolved,
+	})
+	require.NoError(t, err)
+	require.InDelta(t, 300000*10e-6, got.InputCost, 1e-9)
+	require.InDelta(t, 1000*40e-6, got.OutputCost, 1e-9)
+	require.False(t, got.LongContextBillingApplied)
+}
+
+// 目录阶梯跟随分组长上下文开关；开启时为整单换档（输入 ×2、输出 ×1.5）。
+func TestCalculateTokenCostForRequest_CatalogLadderFollowsGroupToggle(t *testing.T) {
+	bs, resolver := newTokenCostTestEnv(t, PlatformGemini, nil, geminiLadderCatalogStub(t))
 	tokens := UsageTokens{InputTokens: 300000, OutputTokens: 1000}
-	rule := bs.LegacyLongContextRule(PlatformGemini)

 	for _, enabled := range []bool{true, false} {
 		group := &Group{ID: 100, Platform: PlatformGemini, LongContextPricingEnabled: enabled}
@@ -89,32 +115,39 @@ func TestCalculateTokenCostForRequest_LegacyRuleFollowsGroupToggle(t *testing.T)

 		got, err := bs.CalculateTokenCostForRequest(TokenCostRequest{
 			Ctx: context.Background(), Model: "gemini-2.5-pro", Group: group, Tokens: tokens, RateMultiplier: 1,
-			Resolver: resolver, Resolved: resolved, LegacyLongContext: rule,
+			Resolver: resolver, Resolved: resolved,
 		})
 		require.NoError(t, err)
 		if enabled {
-			want, err := bs.CalculateCostWithLongContext("gemini-2.5-pro", tokens, 1, rule.Threshold, rule.Multiplier)
-			require.NoError(t, err)
-			require.Equal(t, want, got)
-			// 输入 200K × 1.25e-6 + 超出 100K × 1.25e-6 × 2 = 0.5；输出 1000 × 10e-6 = 0.01。
-			// 旧路径的加倍只体现在 ActualCost（分项 InputCost 不含倍率），探针也据此取值。
-			require.InDelta(t, 0.51, got.ActualCost, 1e-9)
+			// 300K × 1.25e-6 × 2 = 0.75；1000 × 10e-6 × 1.5 = 0.015
+			require.InDelta(t, 0.765, got.ActualCost, 1e-9)
 			require.True(t, got.LongContextBillingApplied)
 		} else {
-			want, err := bs.CalculateCostUnified(CostInput{
-				Ctx: context.Background(), Model: "gemini-2.5-pro", GroupID: &gid, Group: group, Tokens: tokens,
-				RequestCount: 1, RateMultiplier: 1, Resolver: resolver, Resolved: resolved,
-			})
-			require.NoError(t, err)
-			require.Equal(t, want, got)
 			require.InDelta(t, 0.385, got.ActualCost, 1e-9)
 			require.False(t, got.LongContextBillingApplied)
 		}
 	}
 }

+// 目录条目没有阶梯字段时，开关开启也不产生阶梯。
+func TestCalculateTokenCostForRequest_NoLadderFieldsMeansNoLadder(t *testing.T) {
+	bs, resolver := newTokenCostTestEnv(t, PlatformGemini, nil, geminiCatalogStub())
+	group := &Group{ID: 100, Platform: PlatformGemini, LongContextPricingEnabled: true}
+	gid := group.ID
+	resolved := resolver.Resolve(context.Background(), PricingInput{Model: "gemini-2.5-pro", GroupID: &gid, Group: group})
+
+	tokens := UsageTokens{InputTokens: 300000, OutputTokens: 1000}
+	got, err := bs.CalculateTokenCostForRequest(TokenCostRequest{
+		Ctx: context.Background(), Model: "gemini-2.5-pro", Group: group, Tokens: tokens, RateMultiplier: 1,
+		Resolver: resolver, Resolved: resolved,
+	})
+	require.NoError(t, err)
+	require.InDelta(t, 0.385, got.ActualCost, 1e-9)
+	require.False(t, got.LongContextBillingApplied)
+}
+
 func TestCalculateTokenCostForRequest_BuiltInPricingUsesUnifiedPath(t *testing.T) {
-	bs, resolver := newTokenCostTestEnv(t, PlatformOpenAI, nil, nil)
+	bs, resolver := newTokenCostTestEnv(t, PlatformOpenAI, nil, newStubPricingServiceFromJSON(t, openAILadderCatalogJSON))
 	group := &Group{ID: 100, Platform: PlatformOpenAI, LongContextPricingEnabled: true}
 	gid := group.ID
 	tokens := UsageTokens{InputTokens: 300000, OutputTokens: 1000}
@@ -456,45 +456,6 @@ func TestGatewayServiceRecordUsage_UsageLogWriteErrorDoesNotSkipBilling(t *testi
 	require.Equal(t, 1, quotaSvc.quotaCalls)
 }

-func TestGatewayServiceRecordUsageWithLongContext_BillingUsesDetachedContext(t *testing.T) {
-	usageRepo := &openAIRecordUsageLogRepoStub{inserted: false, err: context.DeadlineExceeded}
-	userRepo := &openAIRecordUsageUserRepoStub{}
-	subRepo := &openAIRecordUsageSubRepoStub{}
-	quotaSvc := &openAIRecordUsageAPIKeyQuotaStub{}
-	svc := newGatewayRecordUsageServiceForTest(usageRepo, userRepo, subRepo)
-
-	reqCtx, cancel := context.WithCancel(context.Background())
-	cancel()
-
-	err := svc.RecordUsageWithLongContext(reqCtx, &RecordUsageLongContextInput{
-		Result: &ForwardResult{
-			RequestID: "gateway_long_context_detached_ctx",
-			Usage: ClaudeUsage{
-				InputTokens:  12,
-				OutputTokens: 8,
-			},
-			Model:    "claude-sonnet-4",
-			Duration: time.Second,
-		},
-		APIKey: &APIKey{
-			ID:    502,
-			Quota: 100,
-		},
-		User:                  &User{ID: 602},
-		Account:               &Account{ID: 702},
-		LongContextThreshold:  200000,
-		LongContextMultiplier: 2,
-		APIKeyService:         quotaSvc,
-	})
-
-	require.NoError(t, err)
-	require.Equal(t, 1, usageRepo.calls)
-	require.Equal(t, 1, userRepo.deductCalls)
-	require.NoError(t, userRepo.lastCtxErr)
-	require.Equal(t, 1, quotaSvc.quotaCalls)
-	require.NoError(t, quotaSvc.lastQuotaCtxErr)
-}
-
 func TestGatewayServiceRecordUsage_UsesFallbackRequestIDForUsageLog(t *testing.T) {
 	usageRepo := &openAIRecordUsageLogRepoStub{}
 	userRepo := &openAIRecordUsageUserRepoStub{}
@@ -596,13 +596,6 @@ func writeUsageLogBestEffort(ctx context.Context, repo UsageLogRepository, usage
 	}
 }

-// recordUsageOpts 内部选项，参数化普通计费与长上下文计费的差异点。
-type recordUsageOpts struct {
-	// 长上下文计费（仅 Gemini 路径需要）
-	LongContextThreshold  int
-	LongContextMultiplier float64
-}
-
 // RecordUsage 记录使用量并扣费（或更新订阅用量）
 func (s *GatewayService) RecordUsage(ctx context.Context, input *RecordUsageInput) error {
 	return s.recordUsageCore(ctx, &recordUsageCoreInput{
@@ -622,54 +615,6 @@ func (s *GatewayService) RecordUsage(ctx context.Context, input *RecordUsageInpu
 		APIKeyService:      input.APIKeyService,
 		QuotaPlatform:      input.QuotaPlatform,
 		ChannelUsageFields: input.ChannelUsageFields,
-	}, &recordUsageOpts{})
-}
-
-// RecordUsageLongContextInput 记录使用量的输入参数（支持长上下文双倍计费）
-type RecordUsageLongContextInput struct {
-	Result                *ForwardResult
-	APIKey                *APIKey
-	User                  *User
-	Account               *Account
-	Subscription          *UserSubscription  // 可选：订阅信息
-	PricingAt             time.Time          // token 售价固定时刻；零值保持既有的记录时刻语义
-	InboundEndpoint       string             // 入站端点（客户端请求路径）
-	UpstreamEndpoint      string             // 上游端点（标准化后的上游路径）
-	UserAgent             string             // 请求的 User-Agent
-	IPAddress             string             // 请求的客户端 IP 地址
-	SessionID             string             // 客户端显式会话标识（session_id / X-Session-Id 等请求头），仅用于用量行会话关联
-	RequestPayloadHash    string             // 请求体语义哈希，用于降低 request_id 误复用时的静默误去重风险
-	LongContextThreshold  int                // 长上下文阈值（如 200000）
-	LongContextMultiplier float64            // 超出阈值部分的倍率（如 2.0）
-	ForceCacheBilling     bool               // 强制缓存计费：将 input_tokens 转为 cache_read 计费（用于粘性会话切换）
-	APIKeyService         APIKeyQuotaUpdater // API Key 配额服务（可选）
-	QuotaPlatform         string             // user×platform 配额计量平台：handler 在请求 ctx 内经 QuotaPlatform() 算定后传入（后扣运行在 worker 池 background ctx 上，取不到 ForcePlatform）
-
-	ChannelUsageFields // 渠道映射信息（由 handler 在 Forward 前解析）
-}
-
-// RecordUsageWithLongContext 记录使用量并扣费，支持长上下文双倍计费（用于 Gemini）
-func (s *GatewayService) RecordUsageWithLongContext(ctx context.Context, input *RecordUsageLongContextInput) error {
-	return s.recordUsageCore(ctx, &recordUsageCoreInput{
-		Result:             input.Result,
-		APIKey:             input.APIKey,
-		User:               input.User,
-		Account:            input.Account,
-		Subscription:       input.Subscription,
-		PricingAt:          input.PricingAt,
-		InboundEndpoint:    input.InboundEndpoint,
-		UpstreamEndpoint:   input.UpstreamEndpoint,
-		UserAgent:          input.UserAgent,
-		IPAddress:          input.IPAddress,
-		SessionID:          input.SessionID,
-		RequestPayloadHash: input.RequestPayloadHash,
-		ForceCacheBilling:  input.ForceCacheBilling,
-		APIKeyService:      input.APIKeyService,
-		QuotaPlatform:      input.QuotaPlatform,
-		ChannelUsageFields: input.ChannelUsageFields,
-	}, &recordUsageOpts{
-		LongContextThreshold:  input.LongContextThreshold,
-		LongContextMultiplier: input.LongContextMultiplier,
 	})
 }

@@ -767,9 +712,8 @@ func logResponseModelBillingApplied(component string, account *Account, requestI
 	slog.Info("billing.response_model_applied", attrs...)
 }

-// recordUsageCore 是 RecordUsage 和 RecordUsageWithLongContext 的统一实现。
-// LongContextThreshold > 0 时 Token 计费回退走 CalculateCostWithLongContext。
-func (s *GatewayService) recordUsageCore(ctx context.Context, input *recordUsageCoreInput, opts *recordUsageOpts) error {
+// recordUsageCore 是 RecordUsage 的核心实现。
+func (s *GatewayService) recordUsageCore(ctx context.Context, input *recordUsageCoreInput) error {
 	result := input.Result
 	apiKey := input.APIKey
 	user := input.User
@@ -839,7 +783,7 @@ func (s *GatewayService) recordUsageCore(ctx context.Context, input *recordUsage
 	}

 	// 计算费用
-	cost := s.calculateRecordUsageCost(ctx, result, apiKey, billingModel, multiplier, imageMultiplier, pricingAt, opts)
+	cost := s.calculateRecordUsageCost(ctx, result, apiKey, billingModel, multiplier, imageMultiplier, pricingAt)
 	// response_model：按上游成功响应自报的模型计费（渠道显式开启才生效）。
 	// 采纳条件见 responseModelBillingDeclaration + hasIdentifiedResponseModelPricing
 	// + responseModelBillingAdoptable。任一条件不满足都静默回落基线，即开启本模式前的
@@ -851,7 +795,7 @@ func (s *GatewayService) recordUsageCore(ctx context.Context, input *recordUsage
 		result.ImageCount > 0 || result.AudioUsage != nil || result.SearchCount > 0,
 	); responseModel != "" && !strings.EqualFold(responseModel, strings.TrimSpace(billingModel)) {
 		if identified, responseChannelPriced := s.hasIdentifiedResponseModelPricing(ctx, responseModel, apiKey); identified {
-			responseCost := s.calculateRecordUsageCost(ctx, result, apiKey, responseModel, multiplier, imageMultiplier, pricingAt, opts)
+			responseCost := s.calculateRecordUsageCost(ctx, result, apiKey, responseModel, multiplier, imageMultiplier, pricingAt)
 			baselineChannelPriced := s.resolveChannelPricing(ctx, billingModel, apiKey) != nil
 			if responseModelBillingAdoptable(cost, responseCost, baselineChannelPriced, responseChannelPriced) {
 				// billingModel 到此为止只是定价查表的入参，后续流程只消费 cost，
@@ -872,7 +816,7 @@ func (s *GatewayService) recordUsageCore(ctx context.Context, input *recordUsage
 	// 创建使用日志
 	accountRateMultiplier := account.BillingRateMultiplier()
 	usageLog := s.buildRecordUsageLog(ctx, input, result, apiKey, user, account, subscription,
-		requestedModel, multiplier, imageMultiplier, accountRateMultiplier, billingType, cacheTTLOverridden, cost, opts)
+		requestedModel, multiplier, imageMultiplier, accountRateMultiplier, billingType, cacheTTLOverridden, cost)

 	// 计算账号统计定价费用（使用最终上游模型匹配自定义规则）
 	if apiKey.GroupID != nil {
@@ -932,7 +876,7 @@ func (s *GatewayService) recordUsageCore(ctx context.Context, input *recordUsage
 	return nil
 }

-// calculateRecordUsageCost 根据请求类型和选项计算费用。
+// calculateRecordUsageCost 根据请求类型计算费用。
 func (s *GatewayService) calculateRecordUsageCost(
 	ctx context.Context,
 	result *ForwardResult,
@@ -941,12 +885,11 @@ func (s *GatewayService) calculateRecordUsageCost(
 	multiplier float64,
 	imageMultiplier float64,
 	pricingAt time.Time,
-	opts *recordUsageOpts,
 ) *CostBreakdown {
 	// 图片生成：渠道定价为 token 计费时走 token 路径，否则走图片计费
 	if result.ImageCount > 0 {
 		if resolved := s.resolveChannelPricing(ctx, billingModel, apiKey); resolved != nil && resolved.Mode == BillingModeToken {
-			return s.calculateTokenCost(ctx, result, apiKey, billingModel, multiplier, pricingAt, opts)
+			return s.calculateTokenCost(ctx, result, apiKey, billingModel, multiplier, pricingAt)
 		}
 		return s.calculateImageCost(ctx, result, apiKey, billingModel, imageMultiplier)
 	}
@@ -970,7 +913,7 @@ func (s *GatewayService) calculateRecordUsageCost(
 	}

 	// Token 计费；SearchCount 为叠加 surcharge（不替代 token）。
-	tokenCost := s.calculateTokenCost(ctx, result, apiKey, billingModel, multiplier, pricingAt, opts)
+	tokenCost := s.calculateTokenCost(ctx, result, apiKey, billingModel, multiplier, pricingAt)
 	if result.SearchCount > 0 {
 		price := groupSearchPricePer1kFromAPIKey(apiKey)
 		if price != nil && *price == 0 {
@@ -1122,7 +1065,7 @@ func (s *GatewayService) calculateImageCost(
 	return s.billingService.CalculateImageCost(billingModel, sizeTier, result.ImageCount, groupConfig, multiplier)
 }

-// calculateTokenCost 计算 Token 计费：路径选择（分组/渠道定价 → 旧长上下文规则 → 内置定价）
+// calculateTokenCost 计算 Token 计费：路径选择（分组/渠道定价 → 内置定价）
 // 统一交给 BillingService.CalculateTokenCostForRequest，与模型广场的阶梯表查询同源。
 func (s *GatewayService) calculateTokenCost(
 	ctx context.Context,
@@ -1131,7 +1074,6 @@ func (s *GatewayService) calculateTokenCost(
 	billingModel string,
 	multiplier float64,
 	pricingAt time.Time,
-	opts *recordUsageOpts,
 ) *CostBreakdown {
 	tokens := UsageTokens{
 		InputTokens:           result.Usage.InputTokens,
@@ -1148,22 +1090,17 @@ func (s *GatewayService) calculateTokenCost(
 		gid := apiKey.Group.ID
 		resolved = s.resolver.Resolve(ctx, PricingInput{Model: billingModel, GroupID: &gid, Group: apiKey.Group})
 	}
-	var legacy *LegacyLongContextRule
-	if opts.LongContextThreshold > 0 {
-		legacy = &LegacyLongContextRule{Threshold: opts.LongContextThreshold, Multiplier: opts.LongContextMultiplier}
-	}

 	cost, err := s.billingService.CalculateTokenCostForRequest(TokenCostRequest{
-		Ctx:               ctx,
-		Model:             billingModel,
-		Group:             apiKey.Group,
-		Tokens:            tokens,
-		RateMultiplier:    multiplier,
-		PricingAt:         pricingAt,
-		ServiceTier:       optionalStringValue(result.ServiceTier),
-		Resolver:          s.resolver,
-		Resolved:          resolved,
-		LegacyLongContext: legacy,
+		Ctx:            ctx,
+		Model:          billingModel,
+		Group:          apiKey.Group,
+		Tokens:         tokens,
+		RateMultiplier: multiplier,
+		PricingAt:      pricingAt,
+		ServiceTier:    optionalStringValue(result.ServiceTier),
+		Resolver:       s.resolver,
+		Resolved:       resolved,
 	})
 	if err != nil {
 		logger.LegacyPrintf("service.gateway", "Calculate cost failed: %v", err)
@@ -1172,14 +1109,6 @@ func (s *GatewayService) calculateTokenCost(
 	return cost
 }

-// LegacyLongContextRule 透传 BillingService 的平台旧长上下文规则，供入口 handler 取用。
-func (s *GatewayService) LegacyLongContextRule(platform string) *LegacyLongContextRule {
-	if s == nil || s.billingService == nil {
-		return nil
-	}
-	return s.billingService.LegacyLongContextRule(platform)
-}
-
 // buildRecordUsageLog 构建使用日志并设置计费模式。
 func (s *GatewayService) buildRecordUsageLog(
 	ctx context.Context,
@@ -1196,7 +1125,6 @@ func (s *GatewayService) buildRecordUsageLog(
 	billingType int8,
 	cacheTTLOverridden bool,
 	cost *CostBreakdown,
-	opts *recordUsageOpts,
 ) *UsageLog {
 	durationMs := int(result.Duration.Milliseconds())
 	requestID := resolveUsageBillingRequestID(ctx, result.RequestID)
@@ -362,7 +362,8 @@ func TestListGroups_TokenLadderFollowsGroupToggle(t *testing.T) {
 		{ID: 10, Name: "on", Platform: PlatformOpenAI, RateMultiplier: 1, LongContextPricingEnabled: true},
 		{ID: 20, Name: "off", Platform: PlatformOpenAI, RateMultiplier: 2, LongContextPricingEnabled: false},
 	}
-	svc := newPlazaServiceWithBilling(channels, groups, map[int64]string{10: PlatformOpenAI, 20: PlatformOpenAI}, nil)
+	svc := newPlazaServiceWithBilling(channels, groups, map[int64]string{10: PlatformOpenAI, 20: PlatformOpenAI},
+		newStubPricingServiceFromJSON(t, openAILadderCatalogJSON))
 	out, err := svc.ListGroups(context.Background())
 	require.NoError(t, err)
 	require.Len(t, out, 2)
@@ -395,26 +396,27 @@ func TestListGroups_TokenLadderFollowsGroupToggle(t *testing.T) {
 	}
 }

-func TestListGroups_GeminiLegacyRuleShownAsMarginal(t *testing.T) {
+func TestListGroups_GeminiCatalogLadderShownWholeRequest(t *testing.T) {
 	channels := []Channel{{
 		ID: 1, Name: "ch", Status: StatusActive, GroupIDs: []int64{10},
 		ModelMapping: map[string]map[string]string{PlatformGemini: {"gemini-2.5-pro": "gemini-2.5-pro"}},
 	}}
 	groups := []Group{{ID: 10, Name: "g", Platform: PlatformGemini, RateMultiplier: 1, LongContextPricingEnabled: true}}
-	svc := newPlazaServiceWithBilling(channels, groups, map[int64]string{10: PlatformGemini}, geminiCatalogStub())
+	svc := newPlazaServiceWithBilling(channels, groups, map[int64]string{10: PlatformGemini},
+		newStubPricingServiceFromJSON(t, geminiLadderCatalogJSON))
 	out, err := svc.ListGroups(context.Background())
 	require.NoError(t, err)
 	require.Len(t, out, 1)
 	m := out[0].Models[0]
-	require.Equal(t, ContextPricingBasisMarginal, m.LongContextBasis)
+	require.Equal(t, ContextPricingBasisWholeRequest, m.LongContextBasis)
 	require.Len(t, m.Pricing.Intervals, 2)
 	require.Equal(t, "≤200K", m.Pricing.Intervals[0].TierLabel)
 	require.Equal(t, ">200K", m.Pricing.Intervals[1].TierLabel)
 	require.InDelta(t, 2.5e-6, *m.Pricing.Intervals[1].InputPrice, 1e-15)
-	require.InDelta(t, 10e-6, *m.Pricing.Intervals[1].OutputPrice, 1e-15)
-	// 官方参考不套用站内旧规则
+	require.InDelta(t, 15e-6, *m.Pricing.Intervals[1].OutputPrice, 1e-15)
+	// 官方参考价与实付同源：都来自目录数据的阶梯字段
 	require.NotNil(t, m.OfficialPricing)
-	require.Empty(t, m.OfficialPricing.Intervals)
+	require.Len(t, m.OfficialPricing.Intervals, 2)
 }

 func TestListGroups_GroupTokenCardOverridesChannelPricing(t *testing.T) {
@@ -1182,11 +1182,21 @@ func TestOpenAIGatewayServiceRecordUsage_Gpt54LongContextBillingDisabledWhenGrou
 	require.Equal(t, 1, userRepo.deductCalls)
 }

+// swapInOpenAILadderCatalog 给测试服务换上带 above_272k 阶梯字段的目录：
+// 静态兜底价已不带阶梯，长上下文相关测试需要目录数据。
+func swapInOpenAILadderCatalog(t *testing.T, svc *OpenAIGatewayService) {
+	t.Helper()
+	cfg := &config.Config{}
+	cfg.Default.RateMultiplier = 1.1
+	svc.billingService = NewBillingService(cfg, newStubPricingServiceFromJSON(t, openAILadderCatalogJSON))
+}
+
 func TestOpenAIGatewayServiceRecordUsage_Gpt54LongContextBillingEnabledPerAccount(t *testing.T) {
 	usageRepo := &openAIRecordUsageLogRepoStub{inserted: true}
 	userRepo := &openAIRecordUsageUserRepoStub{}
 	subRepo := &openAIRecordUsageSubRepoStub{}
 	svc := newOpenAIRecordUsageServiceForTest(usageRepo, userRepo, subRepo, nil)
+	swapInOpenAILadderCatalog(t, svc)

 	err := svc.RecordUsage(context.Background(), &OpenAIRecordUsageInput{
 		Result: &OpenAIForwardResult{
@@ -1227,6 +1237,7 @@ func TestOpenAIGatewayServiceRecordUsage_GroupOrAccountLongContextAllows(t *test
 	t.Run("group on account off", func(t *testing.T) {
 		usageRepo := &openAIRecordUsageLogRepoStub{inserted: true}
 		svc := newOpenAIRecordUsageServiceForTest(usageRepo, &openAIRecordUsageUserRepoStub{}, &openAIRecordUsageSubRepoStub{}, nil)
+		swapInOpenAILadderCatalog(t, svc)
 		err := svc.RecordUsage(context.Background(), &OpenAIRecordUsageInput{
 			Result:  &OpenAIForwardResult{RequestID: "resp_and_off", Usage: tokens, Model: "gpt-5.4-2026-03-05", Duration: time.Second},
 			APIKey:  openAIRecordUsageAPIKeyWithGroup(svc, 1020, true),
@@ -1242,6 +1253,7 @@ func TestOpenAIGatewayServiceRecordUsage_GroupOrAccountLongContextAllows(t *test
 	t.Run("group off account on", func(t *testing.T) {
 		usageRepo := &openAIRecordUsageLogRepoStub{inserted: true}
 		svc := newOpenAIRecordUsageServiceForTest(usageRepo, &openAIRecordUsageUserRepoStub{}, &openAIRecordUsageSubRepoStub{}, nil)
+		swapInOpenAILadderCatalog(t, svc)
 		err := svc.RecordUsage(context.Background(), &OpenAIRecordUsageInput{
 			Result: &OpenAIForwardResult{RequestID: "resp_and_group_off", Usage: tokens, Model: "gpt-5.4-2026-03-05", Duration: time.Second},
 			APIKey: openAIRecordUsageAPIKeyWithGroup(svc, 1021, false),
@@ -1260,6 +1272,7 @@ func TestOpenAIGatewayServiceRecordUsage_GroupOrAccountLongContextAllows(t *test
 	t.Run("group on account on", func(t *testing.T) {
 		usageRepo := &openAIRecordUsageLogRepoStub{inserted: true}
 		svc := newOpenAIRecordUsageServiceForTest(usageRepo, &openAIRecordUsageUserRepoStub{}, &openAIRecordUsageSubRepoStub{}, nil)
+		swapInOpenAILadderCatalog(t, svc)
 		err := svc.RecordUsage(context.Background(), &OpenAIRecordUsageInput{
 			Result: &OpenAIForwardResult{RequestID: "resp_and_on", Usage: tokens, Model: "gpt-5.4-2026-03-05", Duration: time.Second},
 			APIKey: openAIRecordUsageAPIKeyWithGroup(svc, 1022, true),
@@ -1352,6 +1365,7 @@ func TestOpenAIGatewayServiceRecordUsage_SparkShadowUsesCurrentParentBillingSett
 				&openAIRecordUsageSubRepoStub{},
 				nil,
 			)
+			swapInOpenAILadderCatalog(t, svc)
 			svc.accountRepo = accountRepo
 			parentID := int64(4016)

@@ -2784,7 +2798,6 @@ func TestGatewayServiceCalculateRecordUsageCost_ChannelImageBillingUsesImageCoun
 		0.15,
 		1.0,
 		time.Time{},
-		nil,
 	)

 	require.NotNil(t, cost)
@@ -2824,7 +2837,6 @@ func TestGatewayServiceCalculateRecordUsageCost_ChannelImageBillingUsesSizeTier(
 		1.0,
 		1.0,
 		time.Time{},
-		nil,
 	)

 	require.NotNil(t, cost)
@@ -2857,7 +2869,6 @@ func TestGatewayServiceCalculateRecordUsageCost_GroupImagePriceOverridesChannelI
 		1.0,
 		1.0,
 		time.Time{},
-		nil,
 	)

 	require.NotNil(t, cost)
@@ -2947,7 +2958,6 @@ func TestGatewayServiceCalculateRecordUsageCost_ChannelImageBillingNormalizesMis
 		1.0,
 		1.0,
 		time.Time{},
-		nil,
 	)

 	require.NotNil(t, cost)
@@ -1,6 +1,7 @@
 package service

 import (
+	"bytes"
 	"context"
 	"crypto/sha256"
 	"encoding/hex"
@@ -10,6 +11,7 @@ import (
 	"path/filepath"
 	"regexp"
 	"sort"
+	"strconv"
 	"strings"
 	"sync"
 	"time"
@@ -22,18 +24,19 @@ import (
 )

 var (
-	openAIModelDatePattern     = regexp.MustCompile(`-\d{8}$`)
-	openAIModelBasePattern     = regexp.MustCompile(`^(gpt-\d+(?:\.\d+)?)(?:-|$)`)
+	openAIModelDatePattern = regexp.MustCompile(`-\d{8}$`)
+	openAIModelBasePattern = regexp.MustCompile(`^(gpt-\d+(?:\.\d+)?)(?:-|$)`)
+	// aboveTierPricePattern 匹配 LiteLLM 长上下文绝对价字段名
+	// （input_cost_per_token_above_272k_tokens / output_cost_per_token_above_200k_tokens 等）。
+	// 带 _flex/_priority 服务档后缀的变体与 cache 侧字段不参与阈值/倍率折算。
+	aboveTierPricePattern      = regexp.MustCompile(`^(input|output)_cost_per_token_above_(\d+)k_tokens$`)
 	openAIGPT54FallbackPricing = &LiteLLMModelPricing{
-		InputCostPerToken:               2.5e-06, // $2.5 per MTok
-		OutputCostPerToken:              1.5e-05, // $15 per MTok
-		CacheReadInputTokenCost:         2.5e-07, // $0.25 per MTok
-		LongContextInputTokenThreshold:  272000,
-		LongContextInputCostMultiplier:  2.0,
-		LongContextOutputCostMultiplier: 1.5,
-		LiteLLMProvider:                 "openai",
-		Mode:                            "chat",
-		SupportsPromptCaching:           true,
+		InputCostPerToken:       2.5e-06, // $2.5 per MTok
+		OutputCostPerToken:      1.5e-05, // $15 per MTok
+		CacheReadInputTokenCost: 2.5e-07, // $0.25 per MTok
+		LiteLLMProvider:         "openai",
+		Mode:                    "chat",
+		SupportsPromptCaching:   true,
 	}
 	openAIGPT56SolFallbackPricing = &LiteLLMModelPricing{
 		InputCostPerToken:                   5e-06,
@@ -44,9 +47,6 @@ var (
 		CacheCreationInputTokenCostPriority: 1.25e-05,
 		CacheReadInputTokenCost:             5e-07,
 		CacheReadInputTokenCostPriority:     1e-06,
-		LongContextInputTokenThreshold:      openAIGPT54LongContextInputThreshold,
-		LongContextInputCostMultiplier:      openAIGPT54LongContextInputMultiplier,
-		LongContextOutputCostMultiplier:     openAIGPT54LongContextOutputMultiplier,
 		SupportsServiceTier:                 true,
 		LiteLLMProvider:                     "openai",
 		Mode:                                "chat",
@@ -61,9 +61,6 @@ var (
 		CacheCreationInputTokenCostPriority: 5e-06,
 		CacheReadInputTokenCost:             2e-07,
 		CacheReadInputTokenCostPriority:     4e-07,
-		LongContextInputTokenThreshold:      openAIGPT54LongContextInputThreshold,
-		LongContextInputCostMultiplier:      openAIGPT54LongContextInputMultiplier,
-		LongContextOutputCostMultiplier:     openAIGPT54LongContextOutputMultiplier,
 		SupportsServiceTier:                 true,
 		LiteLLMProvider:                     "openai",
 		Mode:                                "chat",
@@ -78,9 +75,6 @@ var (
 		CacheCreationInputTokenCostPriority: 5e-07,
 		CacheReadInputTokenCost:             2e-08,
 		CacheReadInputTokenCostPriority:     4e-08,
-		LongContextInputTokenThreshold:      openAIGPT54LongContextInputThreshold,
-		LongContextInputCostMultiplier:      openAIGPT54LongContextInputMultiplier,
-		LongContextOutputCostMultiplier:     openAIGPT54LongContextOutputMultiplier,
 		SupportsServiceTier:                 true,
 		LiteLLMProvider:                     "openai",
 		Mode:                                "chat",
@@ -408,6 +402,7 @@ func (s *PricingService) downloadPricingData() error {

 	// 更新内存数据
 	s.mu.Lock()
+	warnDroppedLongContextLadders(s.pricingData, data)
 	s.pricingData = data
 	s.lastUpdated = time.Now()
 	s.localHash = syncHash
@@ -500,6 +495,13 @@ func (s *PricingService) parsePricingData(body []byte) (map[string]*LiteLLMModel
 			pricing.InputCostPerImageToken = *entry.InputCostPerImageToken
 		}

+		hasExplicitLongContext := entry.LongContextInputTokenThreshold != nil ||
+			entry.LongContextInputCostMultiplier != nil ||
+			entry.LongContextOutputCostMultiplier != nil
+		if !hasExplicitLongContext {
+			deriveLongContextFromAboveTierFields(rawEntry, pricing)
+		}
+
 		result[modelName] = pricing
 	}

@@ -514,6 +516,81 @@ func (s *PricingService) parsePricingData(body []byte) (map[string]*LiteLLMModel
 	return result, nil
 }

+// deriveLongContextFromAboveTierFields 把 LiteLLM 目录的 *_above_XXXk_tokens 绝对价字段
+// 折算成 long_context_* 阈值+倍率（sub2api 计费机制的内部表达）：阈值取自字段名，
+// 倍率 = above 价 ÷ 基础价。条目显式携带任一 long_context_* 字段（含显式 0）时由
+// 调用方跳过折算，以显式配置为准——显式写 threshold=0 或 multiplier=1 均可关闭该
+// 模型的阶梯。多个阈值并存时取最小阈值。
+// cache_read/cache_creation 的 above 档在计费中统一跟随输入倍率，不单独折算；
+// 现网 OpenAI/Claude 条目（含 5m/1h 分档）的 cache above 档均恰为基础价 × 输入倍率，
+// Gemini pro 系存在无基础价的孤儿 cache_creation above 字段（数据缺陷，计费两侧均按 0）。
+func deriveLongContextFromAboveTierFields(rawEntry json.RawMessage, pricing *LiteLLMModelPricing) {
+	if pricing == nil ||
+		pricing.LongContextInputTokenThreshold > 0 ||
+		pricing.LongContextInputCostMultiplier > 0 ||
+		pricing.LongContextOutputCostMultiplier > 0 {
+		return
+	}
+	if !bytes.Contains(rawEntry, []byte("_above_")) {
+		return
+	}
+	var fields map[string]any
+	if err := json.Unmarshal(rawEntry, &fields); err != nil {
+		return
+	}
+	type tierPrices struct{ input, output float64 }
+	tiers := make(map[int]*tierPrices)
+	for key, value := range fields {
+		m := aboveTierPricePattern.FindStringSubmatch(key)
+		if m == nil {
+			continue
+		}
+		price, ok := value.(float64)
+		if !ok || price <= 0 {
+			continue
+		}
+		thousands, err := strconv.Atoi(m[2])
+		if err != nil || thousands <= 0 {
+			continue
+		}
+		threshold := thousands * 1000
+		tp := tiers[threshold]
+		if tp == nil {
+			tp = &tierPrices{}
+			tiers[threshold] = tp
+		}
+		if m[1] == "input" {
+			tp.input = price
+		} else {
+			tp.output = price
+		}
+	}
+	if len(tiers) == 0 {
+		return
+	}
+	threshold := 0
+	for t := range tiers {
+		if threshold == 0 || t < threshold {
+			threshold = t
+		}
+	}
+	tp := tiers[threshold]
+	inputMultiplier, outputMultiplier := 1.0, 1.0
+	if tp.input > 0 && pricing.InputCostPerToken > 0 {
+		inputMultiplier = tp.input / pricing.InputCostPerToken
+	}
+	if tp.output > 0 && pricing.OutputCostPerToken > 0 {
+		outputMultiplier = tp.output / pricing.OutputCostPerToken
+	}
+	// above 价不高于基础价时视为无附加费，不生成阶梯。
+	if inputMultiplier <= 1 && outputMultiplier <= 1 {
+		return
+	}
+	pricing.LongContextInputTokenThreshold = threshold
+	pricing.LongContextInputCostMultiplier = inputMultiplier
+	pricing.LongContextOutputCostMultiplier = outputMultiplier
+}
+
 // loadPricingData 从本地文件加载价格数据
 func (s *PricingService) loadPricingData(filePath string) error {
 	data, err := os.ReadFile(filePath)
@@ -533,6 +610,7 @@ func (s *PricingService) loadPricingData(filePath string) error {
 	hashStr := hex.EncodeToString(hash[:])

 	s.mu.Lock()
+	warnDroppedLongContextLadders(s.pricingData, pricingData)
 	s.pricingData = pricingData
 	s.localHash = hashStr

@@ -579,6 +657,34 @@ func (s *PricingService) mergeFallbackPricingData(data map[string]*LiteLLMModelP
 	return data
 }

+// warnDroppedLongContextLadders 对比新旧目录数据：原本带长上下文阶梯的条目在新数据里
+// 丢失阈值时打 WARN。阶梯已完全数据驱动（无代码兜底），数据源一次误提交就会把阶梯
+// 静默变成基础价少收（07-14~08-21 漏收事故的形态），这里是唯一的哨兵。
+// 调用方需持有 s.mu 写锁。
+func warnDroppedLongContextLadders(old, next map[string]*LiteLLMModelPricing) {
+	if len(old) == 0 {
+		return
+	}
+	var dropped []string
+	for name, prev := range old {
+		if prev == nil || prev.LongContextInputTokenThreshold <= 0 {
+			continue
+		}
+		if cur, ok := next[name]; ok && (cur == nil || cur.LongContextInputTokenThreshold <= 0) {
+			dropped = append(dropped, name)
+		}
+	}
+	if len(dropped) == 0 {
+		return
+	}
+	sort.Strings(dropped)
+	total := len(dropped)
+	if total > 20 {
+		dropped = append(dropped[:20], "...")
+	}
+	logger.LegacyPrintf("service.pricing", "[Pricing] Long-context ladder dropped for %d model(s) after reload: %s (verify catalog/override data if unintended)", total, strings.Join(dropped, ", "))
+}
+
 // useFallbackPricing 使用回退价格文件
 func (s *PricingService) useFallbackPricing() error {
 	fallbackFile := s.cfg.Pricing.FallbackFile
@@ -109,9 +109,8 @@ func TestBillingService_GPT56CacheWritePricingUsesOfficialMultiplier(t *testing.
 			require.NoError(t, err)
 			require.InDelta(t, tt.input*1.25, pricing.CacheCreationPricePerToken, 1e-12)
 			require.InDelta(t, tt.inputPriority*1.25, pricing.CacheCreationPricePerTokenPriority, 1e-12)
-			require.Equal(t, 272000, pricing.LongContextInputThreshold)
-			require.InDelta(t, 2.0, pricing.LongContextInputMultiplier, 1e-12)
-			require.InDelta(t, 1.5, pricing.LongContextOutputMultiplier, 1e-12)
+			// 阶梯由目录数据驱动：条目无 above/long_context 字段时不再由策略强补。
+			require.Zero(t, pricing.LongContextInputThreshold)

 			tokens := UsageTokens{InputTokens: 700, OutputTokens: 50, CacheCreationTokens: 200, CacheReadTokens: 100}
 			standard, err := svc.CalculateCostWithServiceTier(tt.model, tokens, 1, "")
@@ -129,6 +128,32 @@ func TestBillingService_GPT56CacheWritePricingUsesOfficialMultiplier(t *testing.
 	}
 }

+// gpt56LadderCatalogJSON 三个 5.6 模型的目录条目：above_272k 绝对价 + priority 平价，
+// cache_write 缺失由策略按 1.25 倍输入价补齐。
+const gpt56LadderCatalogJSON = `{
+	"gpt-5.6-sol": {"litellm_provider": "openai", "mode": "chat",
+		"input_cost_per_token": 5e-06, "input_cost_per_token_priority": 1e-05,
+		"output_cost_per_token": 3e-05, "output_cost_per_token_priority": 6e-05,
+		"cache_read_input_token_cost": 5e-07, "cache_read_input_token_cost_priority": 1e-06,
+		"input_cost_per_token_above_272k_tokens": 1e-05,
+		"output_cost_per_token_above_272k_tokens": 4.5e-05,
+		"cache_read_input_token_cost_above_272k_tokens": 1e-06},
+	"gpt-5.6-terra": {"litellm_provider": "openai", "mode": "chat",
+		"input_cost_per_token": 2e-06, "input_cost_per_token_priority": 4e-06,
+		"output_cost_per_token": 1.2e-05, "output_cost_per_token_priority": 2.4e-05,
+		"cache_read_input_token_cost": 2e-07, "cache_read_input_token_cost_priority": 4e-07,
+		"input_cost_per_token_above_272k_tokens": 4e-06,
+		"output_cost_per_token_above_272k_tokens": 1.8e-05,
+		"cache_read_input_token_cost_above_272k_tokens": 4e-07},
+	"gpt-5.6-luna": {"litellm_provider": "openai", "mode": "chat",
+		"input_cost_per_token": 2e-07, "input_cost_per_token_priority": 4e-07,
+		"output_cost_per_token": 1.2e-06, "output_cost_per_token_priority": 2.4e-06,
+		"cache_read_input_token_cost": 2e-08, "cache_read_input_token_cost_priority": 4e-08,
+		"input_cost_per_token_above_272k_tokens": 4e-07,
+		"output_cost_per_token_above_272k_tokens": 1.8e-06,
+		"cache_read_input_token_cost_above_272k_tokens": 4e-08}
+}`
+
 func TestBillingService_GPT56UsesLongContextPricingAcrossModelsAndTiers(t *testing.T) {
 	models := []struct {
 		name               string
@@ -157,7 +182,7 @@ func TestBillingService_GPT56UsesLongContextPricingAcrossModelsAndTiers(t *testi
 	for _, model := range models {
 		for _, tier := range tiers {
 			t.Run(model.name+"/"+tier.name, func(t *testing.T) {
-				svc := NewBillingService(&config.Config{}, nil)
+				svc := NewBillingService(&config.Config{}, newStubPricingServiceFromJSON(t, gpt56LadderCatalogJSON))
 				serviceTier := ""
 				if tier.name != "standard" {
 					serviceTier = tier.name
@@ -174,7 +199,7 @@ func TestBillingService_GPT56UsesLongContextPricingAcrossModelsAndTiers(t *testi
 }

 func TestBillingService_GPT56LongContextBoundaryIsExclusive(t *testing.T) {
-	svc := NewBillingService(&config.Config{}, nil)
+	svc := NewBillingService(&config.Config{}, newStubPricingServiceFromJSON(t, gpt56LadderCatalogJSON))
 	tokens := UsageTokens{InputTokens: 100000, CacheCreationTokens: 100000, CacheReadTokens: 72000, OutputTokens: 10}

 	cost, err := svc.CalculateCost("gpt-5.6-sol", tokens, 1)
@@ -284,9 +309,8 @@ func assertGPT56FallbackPricing(t *testing.T, pricing *ModelPricing, input, cach
 	require.InDelta(t, cached, pricing.CacheReadPricePerToken, 1e-12)
 	require.InDelta(t, cacheWrite, pricing.CacheCreationPricePerToken, 1e-12)
 	require.InDelta(t, output, pricing.OutputPricePerToken, 1e-12)
-	require.Equal(t, 272000, pricing.LongContextInputThreshold)
-	require.InDelta(t, 2.0, pricing.LongContextInputMultiplier, 1e-12)
-	require.InDelta(t, 1.5, pricing.LongContextOutputMultiplier, 1e-12)
+	// 静态兜底只兜基础价；阶梯由目录数据（above_272k 折算或显式字段）驱动。
+	require.Zero(t, pricing.LongContextInputThreshold)
 }

 func TestParsePricingData_KeepsImageOnlyPricing(t *testing.T) {
@@ -436,9 +460,8 @@ func TestGetModelPricing_Gpt54UsesStaticFallbackWhenRemoteMissing(t *testing.T)
 	require.InDelta(t, 2.5e-6, got.InputCostPerToken, 1e-12)
 	require.InDelta(t, 1.5e-5, got.OutputCostPerToken, 1e-12)
 	require.InDelta(t, 2.5e-7, got.CacheReadInputTokenCost, 1e-12)
-	require.Equal(t, 272000, got.LongContextInputTokenThreshold)
-	require.InDelta(t, 2.0, got.LongContextInputCostMultiplier, 1e-12)
-	require.InDelta(t, 1.5, got.LongContextOutputCostMultiplier, 1e-12)
+	// 静态兜底只兜基础价，不携带长上下文阶梯（阶梯由目录数据驱动）。
+	require.Zero(t, got.LongContextInputTokenThreshold)
 }

 func TestGetModelPricing_OpenAICompactAliasUsesStaticFallback(t *testing.T) {
@@ -719,3 +742,154 @@ func TestListModelNamesByProvider_EmptyCatalog(t *testing.T) {
 	require.NotNil(t, got)
 	require.Empty(t, got)
 }
+
+// --- above_XXXk 绝对价字段折算为阈值+倍率 ---
+
+func TestParsePricingData_DerivesLongContextFromAboveTierFields(t *testing.T) {
+	svc := &PricingService{}
+	data, err := svc.parsePricingData([]byte(`{
+		"gpt-above": {"litellm_provider": "openai", "mode": "chat",
+			"input_cost_per_token": 5e-06, "output_cost_per_token": 3e-05,
+			"cache_read_input_token_cost": 5e-07,
+			"input_cost_per_token_above_272k_tokens": 1e-05,
+			"output_cost_per_token_above_272k_tokens": 4.5e-05,
+			"cache_read_input_token_cost_above_272k_tokens": 1e-06,
+			"input_cost_per_token_above_272k_tokens_flex": 5e-06,
+			"output_cost_per_token_above_272k_tokens_flex": 2.25e-05},
+		"gemini-above": {"litellm_provider": "vertex_ai-language-models", "mode": "chat",
+			"input_cost_per_token": 1.25e-06, "output_cost_per_token": 1e-05,
+			"input_cost_per_token_above_200k_tokens": 2.5e-06,
+			"output_cost_per_token_above_200k_tokens": 1.5e-05},
+		"explicit-wins": {"litellm_provider": "openai", "mode": "chat",
+			"input_cost_per_token": 5e-06, "output_cost_per_token": 3e-05,
+			"long_context_input_cost_multiplier": 1,
+			"long_context_output_cost_multiplier": 1,
+			"input_cost_per_token_above_272k_tokens": 1e-05,
+			"output_cost_per_token_above_272k_tokens": 4.5e-05},
+		"no-surcharge": {"litellm_provider": "openai", "mode": "chat",
+			"input_cost_per_token": 5e-06, "output_cost_per_token": 3e-05,
+			"input_cost_per_token_above_272k_tokens": 5e-06,
+			"output_cost_per_token_above_272k_tokens": 3e-05},
+		"cache-only-above": {"litellm_provider": "openai", "mode": "chat",
+			"input_cost_per_token": 5e-06, "output_cost_per_token": 3e-05,
+			"cache_read_input_token_cost_above_272k_tokens": 1e-06},
+		"multi-threshold": {"litellm_provider": "openai", "mode": "chat",
+			"input_cost_per_token": 1e-06, "output_cost_per_token": 2e-06,
+			"input_cost_per_token_above_128k_tokens": 2e-06,
+			"input_cost_per_token_above_272k_tokens": 4e-06}
+	}`))
+	require.NoError(t, err)
+
+	openai := data["gpt-above"]
+	require.Equal(t, 272000, openai.LongContextInputTokenThreshold, "阈值取自字段名（_flex 变体不参与）")
+	require.InDelta(t, 2.0, openai.LongContextInputCostMultiplier, 1e-12)
+	require.InDelta(t, 1.5, openai.LongContextOutputCostMultiplier, 1e-12)
+
+	gemini := data["gemini-above"]
+	require.Equal(t, 200000, gemini.LongContextInputTokenThreshold)
+	require.InDelta(t, 2.0, gemini.LongContextInputCostMultiplier, 1e-12)
+	require.InDelta(t, 1.5, gemini.LongContextOutputCostMultiplier, 1e-12)
+
+	explicit := data["explicit-wins"]
+	require.Zero(t, explicit.LongContextInputTokenThreshold, "显式 long_context_* 字段优先，不做折算")
+	require.InDelta(t, 1.0, explicit.LongContextInputCostMultiplier, 1e-12)
+
+	require.Zero(t, data["no-surcharge"].LongContextInputTokenThreshold, "above 价不高于基础价视为无附加费")
+	require.Zero(t, data["cache-only-above"].LongContextInputTokenThreshold, "仅 cache 侧 above 字段不构成阶梯")
+	require.Equal(t, 128000, data["multi-threshold"].LongContextInputTokenThreshold, "多阈值取最小")
+}
+
+func TestGetModelPricing_XAIThresholdInclusive(t *testing.T) {
+	svc := NewBillingService(&config.Config{}, newStubPricingServiceFromJSON(t, `{
+		"grok-4.5": {"litellm_provider": "xai", "mode": "chat",
+			"input_cost_per_token": 2e-06, "output_cost_per_token": 6e-06,
+			"input_cost_per_token_above_200k_tokens": 4e-06,
+			"output_cost_per_token_above_200k_tokens": 1.2e-05}
+	}`))
+	pricing, err := svc.GetModelPricing("grok-4.5")
+	require.NoError(t, err)
+	require.Equal(t, 200000, pricing.LongContextInputThreshold)
+	require.True(t, pricing.LongContextThresholdInclusive, "xAI 阈值语义为达到即进高档")
+}
+
+// F3：显式 long_context 字段以"字段存在"为准——显式 0 也能压住 above 折算，关闭阶梯。
+func TestParsePricingData_ExplicitZeroThresholdDisablesLadder(t *testing.T) {
+	svc := &PricingService{}
+	data, err := svc.parsePricingData([]byte(`{
+		"gpt-5.5": {"litellm_provider": "openai", "mode": "chat",
+			"input_cost_per_token": 5e-06, "output_cost_per_token": 3e-05,
+			"long_context_input_token_threshold": 0,
+			"input_cost_per_token_above_272k_tokens": 1e-05,
+			"output_cost_per_token_above_272k_tokens": 4.5e-05}
+	}`))
+	require.NoError(t, err)
+	require.Zero(t, data["gpt-5.5"].LongContextInputTokenThreshold)
+	require.Zero(t, data["gpt-5.5"].LongContextInputCostMultiplier)
+}
+
+// F1：显式字段只写了一侧倍率时，缺失侧按 1 计而不是乘 0 免费。
+func TestCalculateCost_PartialLongContextMultiplierDefaultsToOne(t *testing.T) {
+	tokens := UsageTokens{InputTokens: 300000, OutputTokens: 1000, CacheReadTokens: 10000}
+
+	t.Run("only input multiplier", func(t *testing.T) {
+		svc := NewBillingService(&config.Config{}, newStubPricingServiceFromJSON(t, `{
+			"partial-in": {"litellm_provider": "openai", "mode": "chat",
+				"input_cost_per_token": 2e-06, "output_cost_per_token": 1e-05,
+				"cache_read_input_token_cost": 2e-07,
+				"long_context_input_token_threshold": 272000,
+				"long_context_input_cost_multiplier": 2.0}
+		}`))
+		cost, err := svc.CalculateCost("partial-in", tokens, 1.0)
+		require.NoError(t, err)
+		require.True(t, cost.LongContextBillingApplied)
+		require.InDelta(t, 300000*2e-6*2, cost.InputCost, 1e-10)
+		require.InDelta(t, 1000*1e-5, cost.OutputCost, 1e-10, "缺失的 output 倍率按 1 计，不得为 0")
+		require.InDelta(t, 10000*2e-7*2, cost.CacheReadCost, 1e-10)
+	})
+
+	t.Run("only output multiplier", func(t *testing.T) {
+		svc := NewBillingService(&config.Config{}, newStubPricingServiceFromJSON(t, `{
+			"partial-out": {"litellm_provider": "openai", "mode": "chat",
+				"input_cost_per_token": 2e-06, "output_cost_per_token": 1e-05,
+				"cache_read_input_token_cost": 2e-07,
+				"long_context_input_token_threshold": 272000,
+				"long_context_output_cost_multiplier": 1.5}
+		}`))
+		cost, err := svc.CalculateCost("partial-out", tokens, 1.0)
+		require.NoError(t, err)
+		require.True(t, cost.LongContextBillingApplied)
+		require.InDelta(t, 300000*2e-6, cost.InputCost, 1e-10, "缺失的 input 倍率按 1 计，不得为 0")
+		require.InDelta(t, 1000*1e-5*1.5, cost.OutputCost, 1e-10)
+		require.InDelta(t, 10000*2e-7, cost.CacheReadCost, 1e-10, "cache_read 跟随 input 倍率，同样按 1 计")
+	})
+}
+
+// 行为声明：目录带 above_200k 的 Claude sonnet 条目同样获得数据驱动的整单阶梯
+// （与 Anthropic 官方 1M 长上下文定价一致），受分组长上下文开关约束。
+func TestCalculateCost_ClaudeSonnetCatalogLadderIsDataDriven(t *testing.T) {
+	svc := NewBillingService(&config.Config{}, newStubPricingServiceFromJSON(t, `{
+		"claude-sonnet-4-5": {"litellm_provider": "anthropic", "mode": "chat",
+			"input_cost_per_token": 3e-06, "output_cost_per_token": 1.5e-05,
+			"cache_read_input_token_cost": 3e-07,
+			"input_cost_per_token_above_200k_tokens": 6e-06,
+			"output_cost_per_token_above_200k_tokens": 2.25e-05,
+			"cache_read_input_token_cost_above_200k_tokens": 6e-07}
+	}`))
+
+	pricing, err := svc.GetModelPricing("claude-sonnet-4-5")
+	require.NoError(t, err)
+	require.Equal(t, 200000, pricing.LongContextInputThreshold)
+	require.False(t, pricing.LongContextThresholdInclusive, "anthropic 为严格大于")
+
+	over := UsageTokens{InputTokens: 250000, OutputTokens: 1000}
+	cost, err := svc.CalculateCost("claude-sonnet-4-5", over, 1.0)
+	require.NoError(t, err)
+	require.True(t, cost.LongContextBillingApplied)
+	require.InDelta(t, 250000*3e-6*2, cost.InputCost, 1e-10)
+	require.InDelta(t, 1000*1.5e-5*1.5, cost.OutputCost, 1e-10)
+
+	under := UsageTokens{InputTokens: 200000, OutputTokens: 1000}
+	cost, err = svc.CalculateCost("claude-sonnet-4-5", under, 1.0)
+	require.NoError(t, err)
+	require.False(t, cost.LongContextBillingApplied, "恰好 200000 不进高档（严格大于）")
+}
@@ -0,0 +1,33 @@
+package service
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+// openAILadderCatalogJSON 镜像真实同步目录的形态：长上下文用 above_272k 绝对价字段表达，
+// 由解析层折算成阈值+倍率。静态 Go 兜底价不再携带阶梯，阶梯计费一律走目录数据。
+const openAILadderCatalogJSON = `{
+	"gpt-5.4": {"litellm_provider": "openai", "mode": "chat",
+		"input_cost_per_token": 2.5e-06, "output_cost_per_token": 1.5e-05,
+		"cache_read_input_token_cost": 2.5e-07, "cache_creation_input_token_cost": 2.5e-06,
+		"input_cost_per_token_above_272k_tokens": 5e-06,
+		"output_cost_per_token_above_272k_tokens": 2.25e-05,
+		"cache_read_input_token_cost_above_272k_tokens": 5e-07},
+	"gpt-5.5-pro": {"litellm_provider": "openai", "mode": "chat",
+		"input_cost_per_token": 3e-05, "output_cost_per_token": 1.8e-04,
+		"input_cost_per_token_above_272k_tokens": 6e-05,
+		"output_cost_per_token_above_272k_tokens": 2.7e-04}
+}`
+
+// newStubPricingServiceFromJSON 用与生产一致的解析路径（含 above_XXXk 阶梯折算）
+// 从原始目录 JSON 构造目录 stub。无 build tag：带 unit 标签与默认构建的测试文件都会用到。
+func newStubPricingServiceFromJSON(t *testing.T, body string) *PricingService {
+	t.Helper()
+	s := &PricingService{}
+	data, err := s.parsePricingData([]byte(body))
+	require.NoError(t, err)
+	s.pricingData = data
+	return s
+}
@@ -1054,8 +1054,8 @@ plugins:
 # 定价数据源（可选）
 # =============================================================================
 pricing:
-  # URL to fetch model pricing data (default: pinned model-price-repo commit)
-  # 获取模型定价数据的 URL（默认：固定 commit 的 model-price-repo）
+  # URL to fetch model pricing data (default: model-price-repo main branch)
+  # 获取模型定价数据的 URL（默认：model-price-repo main 分支）
  remote_url: "https://raw.githubusercontent.com/Wei-Shaw/model-price-repo/refs/heads/main//model_prices_and_context_window.json"
  # Hash verification URL (optional)
  # 哈希校验 URL（可选）