feat(pricing): add models.dev-backed tier-aware costs

README.md

@@ -77,7 +77,7 @@ See [Choosing A CPA Panel](https://seakee.github.io/CPA-Manager-Plus/docs/en/gui
 
 - Break down calls, tokens, cost, latency, and failures by model, provider, account, credential, API key, project, channel, and time range.
 - Track input, output, reasoning, cache, service tier, and long-context pricing semantics.
-- Sync model prices from LiteLLM and OpenRouter, with local overrides for aliases or internal models.
+- Sync model prices from models.dev first, with LiteLLM and OpenRouter fallbacks plus local overrides for aliases or internal models.
 - Open the [Usage Analytics Demo](https://seakee.github.io/CPA-Manager-Plus/#/demo/usage-analytics).
 
 ### Account Health, Quota, And Automation

apps/docs/en/manual/model-prices.md

@@ -11,12 +11,24 @@ Open the [Model Prices Demo](https://seakee.github.io/CPA-Manager-Plus/#/demo/mo
 
 ## Price Sources
 
-- Public metadata actively synchronized from LiteLLM or OpenRouter.
+- Public metadata synchronized from models.dev first, with LiteLLM and OpenRouter used as fallbacks when the preferred source is unavailable or lacks a model.
 - Local prices added or overridden by the user.
 - Entries for aliases, internal names, or provider-specific variants.
 
 Synchronization only occurs when the user triggers it and may use the current Manager Server proxy configuration.
 
+The same models.dev model ID may be offered by multiple providers, and a real model ID may itself contain `/`. CPAMP compares source identities separately from original model IDs and only matches automatically when the complete pricing metadata, including tiers and experimental-mode prices, is consistent. Identity or price conflicts remain in the candidate-confirmation flow; a same-named LiteLLM or OpenRouter entry cannot bypass that conflict.
+
+The current sync maps models.dev `cost.input`, `cost.output`, `cost.cache_read`, and `cost.cache_write`, converts valid `cost.tiers` context tiers into CPAMP billing rules, and maps `experimental.modes.fast.cost` to short-context Fast/Priority prices. The complete model object remains available in raw metadata; reasoning prices, unknown experimental modes, unknown tier types, and rules that cannot be validated safely do not activate automatic billing.
+
+### Sync failures and last-known-good prices
+
+- When models.dev is temporarily unavailable, CPAMP continues with LiteLLM and OpenRouter.
+- A transient models.dev failure cannot automatically replace a stored models.dev price with a lower-priority source; fallback sources may still fill models that have no local price.
+- A fallback source may replace a model normally when models.dev responds successfully but does not contain that model.
+- If every source fails, synchronization stops before any database write and existing prices remain unchanged.
+- A synchronized price remains the last-known-good value until a later successful sync or a manual edit; `syncedAtMs` indicates its freshness.
+
 ## Supported Billing Semantics
 
 A price rule may include:
@@ -26,11 +38,29 @@ A price rule may include:
 - Cache read, cache write, and cache creation.
 - Fixed per-request cost.
 - `service_tier` differences.
+- models.dev context-price tiers.
 - Long-context thresholds and multipliers.
 - Model alias and billing-model mapping.
 
 Models such as GPT-5.6 may vary by context length, service tier, and cache type. CPAMP can only apply a rule when both the request event and price entry contain the required fields.
 
+### Context-tier semantics
+
+- A tier matches only when normalized input tokens are **strictly greater than** `tier.size`; an exact-threshold request stays in the lower band.
+- When multiple tiers match, CPAMP selects the highest matching threshold.
+- The selected tier's rates apply to the entire request, not only to tokens above the threshold.
+- Input, output, or cache rates omitted by a tier inherit the base price; an explicit zero from models.dev remains zero.
+- CPAMP currently activates only safely validated tiers with `tier.type = context` and a positive threshold. Other rules remain in raw metadata for inspection.
+
+### Fast/Priority semantics
+
+- `experimental.modes.fast.cost` matches both `fast` usage telemetry and API `priority` telemetry.
+- Short-context requests prefer explicit Fast/Priority prices. Missing fields inherit base rates, while explicit zeros remain zero.
+- A matched context tier or the legacy GPT long-context rule uses its standard context price without stacking Fast/Priority pricing.
+- Non-models.dev entries, older data, and models without an explicit mode price retain the existing multiplier as a compatibility fallback.
+
+Model Prices displays synchronized context tiers and service-tier prices as read-only rules. The current manual editor manages base prices only; saving a manual price explicitly clears existing synchronized advanced rules, with a warning shown before saving.
+
 ## Matching Model Names
 
 The client model, CPA alias, provider model, and price-table name may differ. When cost is missing:
@@ -48,5 +78,5 @@ The page uses a compact model-usage summary to show which prices are active. It
 
 - Provider billing remains authoritative.
 - Missing token, service tier, long-context, or cache fields reduce estimate accuracy.
-- Subscriptions, grants, tiered prices, and multiple currencies may not fit a single price entry.
+- Subscriptions, grants, non-context tiers, unsupported dynamic-mode prices, and multiple currencies may not fit a single price entry.
 - Historical cost may be displayed using current prices after an update; the price table is not an immutable billing snapshot.

apps/docs/manual/model-prices.md

@@ -11,12 +11,24 @@ description: 配置 CPA Manager Plus 模型价格、service tier、长上下文
 
 ## 价格来源
 
-- 从 LiteLLM 或 OpenRouter 主动同步的公开元数据。
+- 首选从 models.dev 主动同步的公开元数据；当该来源不可用或缺少模型时，再使用 LiteLLM 和 OpenRouter 回退。
 - 用户手动添加或覆盖的本地价格。
 - 为模型别名、内部名称或 Provider 特定变体维护的条目。
 
 同步只在用户主动触发时发生，可能使用当前 Manager Server 代理设置。
 
+models.dev 中同一个模型 ID 可能由多个 Provider 提供，而且真实模型 ID 本身也可能包含 `/`。CPAMP 会分别比较来源身份和原始模型 ID；只有完整价格元数据（包括阶梯和实验模式价格）明确一致时才自动匹配。任何身份或价格冲突都会进入候选确认流程，LiteLLM 或 OpenRouter 的同名条目不会绕过该冲突。
+
+当前同步会映射 models.dev 的 `cost.input`、`cost.output`、`cost.cache_read` 和 `cost.cache_write`，将有效的 `cost.tiers` 上下文阶梯转换为 CPAMP 计费规则，并将 `experimental.modes.fast.cost` 映射为 Fast/Priority 短上下文价格。完整模型对象仍保存在原始元数据中；reasoning、未知实验模式、未知阶梯类型或无法安全验证的规则不会激活自动计费。
+
+### 同步失败与最后有效价格
+
+- models.dev 暂时不可用时，CPAMP 会继续尝试 LiteLLM 和 OpenRouter。
+- 已保存的 models.dev 价格不会因为本次网络失败而被低优先级来源自动覆盖；回退来源仍可补充本地没有价格的模型。
+- 只有 models.dev 成功响应但明确不包含某个模型时，才允许回退来源正常替换该模型。
+- 如果所有来源都失败，同步在写入数据库前终止，现有价格保持不变。
+- 同步价格会一直作为最后有效数据使用，直到后续成功同步或用户手动修改；`syncedAtMs` 可用于判断数据新鲜度。
+
 ## 当前支持的计费语义
 
 价格结构可能包括：
@@ -26,11 +38,29 @@ description: 配置 CPA Manager Plus 模型价格、service tier、长上下文
 - Cache read、cache write 和 cache creation。
 - 请求级固定费用。
 - `service_tier` 差异。
+- models.dev 上下文价格阶梯。
 - 长上下文阈值和倍率。
 - 模型别名与 billing model 映射。
 
 例如 GPT-5.6 及类似模型可能根据上下文长度、service tier 和缓存类型采用不同价格。只有请求事件带有对应字段且价格规则存在时，CPAMP 才能正确计算。
 
+### 上下文阶梯语义
+
+- 阶梯条件是标准化输入 Token **严格大于** `tier.size`；恰好等于阈值时仍使用较低一档。
+- 如果多个阶梯都满足条件，选择阈值最高的一档。
+- 选中阶梯的费率应用于整个请求，而不是只计算超过阈值的 Token。
+- 阶梯中缺失的输入、输出或缓存费率继承基础价格；models.dev 明确提供的零价格会保留为零。
+- 目前只自动启用 `tier.type = context`、正数阈值且价格结构可安全验证的阶梯。其他规则仍保留在原始元数据中供排查。
+
+### Fast/Priority 语义
+
+- `experimental.modes.fast.cost` 同时匹配使用数据中的 `fast` 和 API `priority`。
+- 短上下文优先使用显式 Fast/Priority 价格；缺失字段继承基础价格，显式零值保持为零。
+- 命中上下文阶梯或旧版 GPT 长上下文规则时，使用对应的标准上下文价格，不再叠加 Fast/Priority。
+- 非 models.dev、旧数据或没有显式模式价格的模型继续使用现有倍率作为兼容回退。
+
+模型价格页只读展示已同步的上下文阶梯和服务层级价格。当前手动编辑器只维护基础价格；保存手动价格会明确清除该模型已有的同步高级规则，界面会在保存前提示。
+
 ## 模型名称匹配
 
 客户端请求名、CPA 路由别名、Provider 实际模型名和价格表名称可能不同。排查成本为空时：
@@ -48,5 +78,5 @@ description: 配置 CPA Manager Plus 模型价格、service tier、长上下文
 
 - Provider 账单是最终依据。
 - 缺失 Token、service tier、长上下文或缓存字段会降低估算精度。
-- 包月、赠送额度、阶梯价和多币种不一定能由单一价格条目完整表达。
+- 包月、赠送额度、非上下文阶梯、未支持的动态模式价格和多币种不一定能由单一价格条目完整表达。
 - 更新价格后，历史成本可能按当前价格重新展示；价格表不是不可变账单快照。

apps/manager-server/cmd/cpa-manager-plus/main.go

@@ -103,13 +103,17 @@ func runServer() {
 	accountActionWorker := worker.NewAccountActionCandidateWorker(db, runtimeSettings.AccountActionsAutoDisable)
 	accountHistoryRollupWorker := worker.NewAccountHistoryRollupWorker(db)
 	accountHistoryRollupWorker.Start(ctx)
+	usagePricingRollupWorker := worker.NewUsagePricingRollupWorker(db)
+	usagePricingRollupWorker.Start(ctx)
+	serverApp.AppContext().ModelPriceService.SetPricesChangedNotifier(usagePricingRollupWorker.Wake)
 	var usageHourlyAggregateWorker *worker.UsageHourlyAggregateWorker
 	if cfg.DashboardHourlyRollupEnabled {
 		usageHourlyAggregateWorker = worker.NewUsageHourlyAggregateWorker(db)
 		usageHourlyAggregateWorker.Start(ctx)
 	}
 	serverApp.AppContext().UsageService.SetEventsInsertedNotifier(func() {
 		accountHistoryRollupWorker.Wake()
+		usagePricingRollupWorker.Wake()
 		if usageHourlyAggregateWorker != nil {
 			usageHourlyAggregateWorker.Wake()
 		}
@@ -125,6 +129,7 @@ func runServer() {
 	manager.SetUsageEventHandler(worker.NewUsageEventFanout(
 		automationRuntime.UsageEventHandler(),
 		accountHistoryRollupWorker,
+		usagePricingRollupWorker,
 		usageHourlyAggregateWorker,
 	))
 
@@ -167,6 +172,7 @@ func runServer() {
 	usageCacheAccountingMigrationWorker := worker.NewUsageCacheAccountingMigrationWorker(db, func() {
 		go runUsageResponseMetadataBackfill(ctx, db)
 		accountHistoryRollupWorker.Wake()
+		usagePricingRollupWorker.Wake()
 		if usageHourlyAggregateWorker != nil {
 			usageHourlyAggregateWorker.Wake()
 		}

apps/manager-server/internal/app/app.go

@@ -15,6 +15,7 @@ import (
 
 type Options struct {
 	EmbeddedPanel               fs.FS
+	ModelsDevModelPriceSyncURL  *string
 	ModelPriceSyncURL           *string
 	OpenRouterModelPriceSyncURL *string
 	ServiceID                   string
@@ -51,12 +52,13 @@ func New(ctx context.Context, cfg config.Config, options Options) (*Context, err
 	if startedAt <= 0 {
 		startedAt = time.Now().UnixMilli()
 	}
-	appCtx := FromExisting(
+	appCtx := FromExistingWithModelsDev(
 		cfg,
 		st,
 		manager,
 		startedAt,
 		options.EmbeddedPanel,
+		options.ModelsDevModelPriceSyncURL,
 		options.ModelPriceSyncURL,
 		options.OpenRouterModelPriceSyncURL,
 		serviceID,

apps/manager-server/internal/app/context.go

@@ -66,6 +66,58 @@ func FromExisting(
 	openRouterModelPriceSyncURL *string,
 	serviceID string,
 	automationRuntimeService ...AutomationRuntimeService,
+) *Context {
+	return fromExisting(
+		cfg,
+		st,
+		collectorManager,
+		startedAt,
+		embeddedPanel,
+		nil,
+		modelPriceSyncURL,
+		openRouterModelPriceSyncURL,
+		serviceID,
+		automationRuntimeService...,
+	)
+}
+
+func FromExistingWithModelsDev(
+	cfg config.Config,
+	st *store.Store,
+	collectorManager *collector.Manager,
+	startedAt int64,
+	embeddedPanel fs.FS,
+	modelsDevModelPriceSyncURL *string,
+	modelPriceSyncURL *string,
+	openRouterModelPriceSyncURL *string,
+	serviceID string,
+	automationRuntimeService ...AutomationRuntimeService,
+) *Context {
+	return fromExisting(
+		cfg,
+		st,
+		collectorManager,
+		startedAt,
+		embeddedPanel,
+		modelsDevModelPriceSyncURL,
+		modelPriceSyncURL,
+		openRouterModelPriceSyncURL,
+		serviceID,
+		automationRuntimeService...,
+	)
+}
+
+func fromExisting(
+	cfg config.Config,
+	st *store.Store,
+	collectorManager *collector.Manager,
+	startedAt int64,
+	embeddedPanel fs.FS,
+	modelsDevModelPriceSyncURL *string,
+	modelPriceSyncURL *string,
+	openRouterModelPriceSyncURL *string,
+	serviceID string,
+	automationRuntimeService ...AutomationRuntimeService,
 ) *Context {
 	var runtimeService AutomationRuntimeService
 	if len(automationRuntimeService) > 0 {
@@ -99,7 +151,7 @@ func FromExisting(
 		DashboardService:               dashboardsvc.New(st, cfg.DashboardHourlyRollupEnabled),
 		CodexInspectionService:         codexinspectionsvc.New(st, managerConfigService),
 		MonitoringService:              monitoringsvc.New(st, cfg.DashboardHourlyRollupEnabled),
-		ModelPriceService:              modelpricesvc.NewMultiSource(st, modelPriceSyncURL, openRouterModelPriceSyncURL, managerConfigService),
+		ModelPriceService:              modelpricesvc.NewMultiSourceWithModelsDev(st, modelsDevModelPriceSyncURL, modelPriceSyncURL, openRouterModelPriceSyncURL, managerConfigService),
 		APIKeyAliasService:             apikeyaliassvc.New(st),
 		AccountActionService:           accountactionsvc.New(st, managerConfigService),
 		AccountProcessingPolicyService: accountProcessingPolicyService,

apps/manager-server/internal/httpapi/server.go

@@ -17,6 +17,7 @@ var embeddedPanel embed.FS
 
 const serviceID = "cpa-manager-plus"
 
+var modelsDevModelPriceSyncURL = "https://models.dev/api.json"
 var modelPriceSyncURL = "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json"
 var openRouterModelPriceSyncURL = "https://openrouter.ai/api/v1/models"
 
@@ -27,12 +28,13 @@ type Server struct {
 
 func New(cfg config.Config, store *store.Store, collector *collector.Manager, automationRuntimeService ...app.AutomationRuntimeService) *Server {
 	startedAt := time.Now().UnixMilli()
-	appCtx := app.FromExisting(
+	appCtx := app.FromExistingWithModelsDev(
 		cfg,
 		store,
 		collector,
 		startedAt,
 		embeddedPanel,
+		&modelsDevModelPriceSyncURL,
 		&modelPriceSyncURL,
 		&openRouterModelPriceSyncURL,
 		serviceID,