@llmops/core 0.6.1 → 0.6.3-beta.1

package/dist/index.cjs

@@ -1,5 +1,5 @@
-const require_db = require('./db-C6ApWDjW.cjs');
-const require_neon_dialect = require('./neon-dialect-SqAJhPFS.cjs');
+const require_db = require('./db-i0OOYxJm.cjs');
+const require_neon_dialect = require('./neon-dialect-oh8u9vRy.cjs');
 let __llmops_gateway = require("@llmops/gateway");
 __llmops_gateway = require_db.__toESM(__llmops_gateway);
 let kysely = require("kysely");
@@ -2004,7 +2004,9 @@ const insertLLMRequestSchema = require_db.zod_default.object({
 	completionTokens: require_db.zod_default.number().int().default(0),
 	totalTokens: require_db.zod_default.number().int().default(0),
 	cachedTokens: require_db.zod_default.number().int().default(0),
+	cacheCreationTokens: require_db.zod_default.number().int().default(0),
 	cost: require_db.zod_default.number().int().default(0),
+	cacheSavings: require_db.zod_default.number().int().default(0),
 	inputCost: require_db.zod_default.number().int().default(0),
 	outputCost: require_db.zod_default.number().int().default(0),
 	endpoint: require_db.zod_default.string(),
@@ -2099,7 +2101,9 @@ const createLLMRequestsDataLayer = (db) => {
 				completionTokens: req.completionTokens,
 				totalTokens: req.totalTokens,
 				cachedTokens: req.cachedTokens,
+				cacheCreationTokens: req.cacheCreationTokens,
 				cost: req.cost,
+				cacheSavings: req.cacheSavings,
 				inputCost: req.inputCost,
 				outputCost: req.outputCost,
 				endpoint: req.endpoint,
@@ -2137,7 +2141,9 @@ const createLLMRequestsDataLayer = (db) => {
 				completionTokens: req.completionTokens,
 				totalTokens: req.totalTokens,
 				cachedTokens: req.cachedTokens,
+				cacheCreationTokens: req.cacheCreationTokens,
 				cost: req.cost,
+				cacheSavings: req.cacheSavings,
 				inputCost: req.inputCost,
 				outputCost: req.outputCost,
 				endpoint: req.endpoint,
@@ -2199,6 +2205,8 @@ const createLLMRequestsDataLayer = (db) => {
 				kysely.sql`COALESCE(SUM(${col$1("promptTokens")}), 0)`.as("totalPromptTokens"),
 				kysely.sql`COALESCE(SUM(${col$1("completionTokens")}), 0)`.as("totalCompletionTokens"),
 				kysely.sql`COALESCE(SUM(${col$1("totalTokens")}), 0)`.as("totalTokens"),
+				kysely.sql`COALESCE(SUM(${col$1("cachedTokens")}), 0)`.as("totalCachedTokens"),
+				kysely.sql`COALESCE(SUM(${col$1("cacheSavings")}), 0)`.as("totalCacheSavings"),
 				kysely.sql`COUNT(*)`.as("requestCount")
 			]).where(kysely.sql`${col$1("createdAt")} >= ${startDate.toISOString()}`).where(kysely.sql`${col$1("createdAt")} <= ${endDate.toISOString()}`);
 			if (configId) query = query.where("configId", "=", configId);
@@ -3712,7 +3720,64 @@ function calculateCost(usage, pricing) {
 	return {
 		inputCost,
 		outputCost,
-		totalCost: inputCost + outputCost
+		totalCost: inputCost + outputCost,
+		cacheSavings: 0
+	};
+}
+/**
+* Get default cache read rate as a fraction of input cost per provider.
+* Used when models.dev doesn't provide cache pricing.
+*/
+function getDefaultCacheReadRate(provider, inputCostPer1M) {
+	switch (provider?.toLowerCase()) {
+		case "anthropic": return inputCostPer1M * .1;
+		case "openai":
+		case "azure-openai": return inputCostPer1M * .5;
+		case "google":
+		case "gemini":
+		case "vertex_ai": return inputCostPer1M * .25;
+		default: return inputCostPer1M * .5;
+	}
+}
+/**
+* Get default cache write/creation rate as a fraction of input cost per provider.
+* Used when models.dev doesn't provide cache pricing.
+*/
+function getDefaultCacheWriteRate(provider, inputCostPer1M) {
+	switch (provider?.toLowerCase()) {
+		case "anthropic": return inputCostPer1M * 1.25;
+		default: return inputCostPer1M;
+	}
+}
+/**
+* Calculate cache-aware cost of an LLM request in micro-dollars.
+*
+* Splits input tokens into uncached, cache-read, and cache-creation buckets,
+* each priced at different rates. Falls back to provider-specific multipliers
+* when models.dev doesn't provide cache pricing.
+*
+* @param usage - Token usage data (with cachedTokens and cacheCreationTokens)
+* @param pricing - Model pricing (may include cacheReadCostPer1M / cacheWriteCostPer1M)
+* @param provider - Provider name for fallback rate selection
+* @returns Cost breakdown in micro-dollars
+*/
+function calculateCacheAwareCost(usage, pricing, provider) {
+	const cachedTokens = usage.cachedTokens ?? 0;
+	const cacheCreationTokens = usage.cacheCreationTokens ?? 0;
+	if (cachedTokens === 0 && cacheCreationTokens === 0) return calculateCost(usage, pricing);
+	const cacheReadRate = pricing.cacheReadCostPer1M ?? getDefaultCacheReadRate(provider, pricing.inputCostPer1M);
+	const cacheWriteRate = pricing.cacheWriteCostPer1M ?? getDefaultCacheWriteRate(provider, pricing.inputCostPer1M);
+	const uncachedInputTokens = Math.max(0, usage.promptTokens - cachedTokens - cacheCreationTokens);
+	const regularInputCost = Math.round(uncachedInputTokens * pricing.inputCostPer1M);
+	const cacheReadCost = Math.round(cachedTokens * cacheReadRate);
+	const cacheWriteCost = Math.round(cacheCreationTokens * cacheWriteRate);
+	const outputCost = Math.round(usage.completionTokens * pricing.outputCostPer1M);
+	const inputCost = regularInputCost + cacheReadCost + cacheWriteCost;
+	return {
+		inputCost,
+		outputCost,
+		totalCost: inputCost + outputCost,
+		cacheSavings: Math.round((cachedTokens + cacheCreationTokens) * pricing.inputCostPer1M) - cacheReadCost - cacheWriteCost
 	};
 }
 /**
@@ -3764,111 +3829,108 @@ function formatCost(microDollars, decimals = 6) {
 
 //#endregion
 //#region src/pricing/provider.ts
-const MODELS_DEV_API = "https://models.dev/api.json";
+const LLMOPS_MODELS_API = "https://models.llmops.build";
+/**
+* Convert price from USD cents per token to dollars per 1M tokens.
+*
+* API returns cents/token. Our system uses dollars/1M tokens.
+* Formula: (centsPerToken / 100) * 1_000_000 = centsPerToken * 10_000
+*/
+function centsPerTokenToCostPer1M(centsPerToken) {
+	return centsPerToken * 1e4;
+}
 /**
-* Pricing provider that fetches data from models.dev API
+* Pricing provider that fetches per-model data from the LLMOps Models API.
 *
 * Features:
-* - Caches pricing data with configurable TTL (default 5 minutes)
-* - Supports fallback to local cache on fetch failure
-* - Thread-safe cache refresh
+* - Per-model in-memory cache with configurable TTL (default 5 minutes)
+* - Deduplicates concurrent fetches for the same model
+* - Caches null results (404s) to avoid repeated lookups
+* - Falls back to stale cache on fetch errors
 */
-var ModelsDevPricingProvider = class {
+var LLMOpsPricingProvider = class {
 	cache = /* @__PURE__ */ new Map();
-	lastFetch = 0;
+	pendingFetches = /* @__PURE__ */ new Map();
 	cacheTTL;
-	fetchPromise = null;
-	ready = false;
-	/**
-	* Create a new ModelsDevPricingProvider
-	*
-	* @param cacheTTL - Cache TTL in milliseconds (default: 5 minutes)
-	*/
-	constructor(cacheTTL = 300 * 1e3) {
-		this.cacheTTL = cacheTTL;
+	baseUrl;
+	constructor(options) {
+		this.cacheTTL = options?.cacheTTL ?? 300 * 1e3;
+		this.baseUrl = options?.baseUrl ?? LLMOPS_MODELS_API;
 	}
-	/**
-	* Generate a cache key for a provider/model combination
-	*/
 	getCacheKey(provider, model) {
 		return `${provider.toLowerCase()}:${model.toLowerCase()}`;
 	}
 	/**
-	* Fetch pricing data from models.dev API
+	* Fetch pricing for a single model from the API
 	*/
-	async fetchPricingData() {
+	async fetchModelPricing(provider, model) {
+		const url = `${this.baseUrl}/model-configs/pricing/${encodeURIComponent(provider)}/${model}`;
 		try {
-			require_db.logger.debug("[Pricing] Fetching pricing data from models.dev");
-			const response = await fetch(MODELS_DEV_API);
-			if (!response.ok) throw new Error(`Failed to fetch models.dev API: ${response.status}`);
-			const data = await response.json();
-			this.cache.clear();
-			for (const [providerId, provider] of Object.entries(data)) {
-				if (!provider.models) continue;
-				for (const [_modelId, model] of Object.entries(provider.models)) {
-					if (!model.cost) continue;
-					const cacheKey = this.getCacheKey(providerId, model.id);
-					this.cache.set(cacheKey, {
-						inputCostPer1M: model.cost.input ?? 0,
-						outputCostPer1M: model.cost.output ?? 0,
-						cacheReadCostPer1M: model.cost.cache_read,
-						cacheWriteCostPer1M: model.cost.cache_write,
-						reasoningCostPer1M: model.cost.reasoning
-					});
-					const nameKey = this.getCacheKey(providerId, model.name);
-					if (nameKey !== cacheKey) this.cache.set(nameKey, this.cache.get(cacheKey));
-				}
+			require_db.logger.debug(`[Pricing] GET ${url}`);
+			const startTime = Date.now();
+			const response = await fetch(url);
+			const elapsed = Date.now() - startTime;
+			require_db.logger.debug(`[Pricing] GET ${url} -> ${response.status} (${elapsed}ms)`);
+			if (response.status === 404) {
+				require_db.logger.debug(`[Pricing] No pricing found for ${provider}/${model}`);
+				return null;
 			}
-			this.lastFetch = Date.now();
-			this.ready = true;
-			require_db.logger.debug(`[Pricing] Cached pricing for ${this.cache.size} models from models.dev`);
+			if (!response.ok) throw new Error(`API returned ${response.status}`);
+			const data = await response.json();
+			if (!data.pay_as_you_go) return null;
+			const payg = data.pay_as_you_go;
+			const pricing = {
+				inputCostPer1M: centsPerTokenToCostPer1M(payg.request_token?.price ?? 0),
+				outputCostPer1M: centsPerTokenToCostPer1M(payg.response_token?.price ?? 0),
+				cacheReadCostPer1M: payg.cache_read_input_token?.price != null ? centsPerTokenToCostPer1M(payg.cache_read_input_token.price) : void 0,
+				cacheWriteCostPer1M: payg.cache_write_input_token?.price != null ? centsPerTokenToCostPer1M(payg.cache_write_input_token.price) : void 0
+			};
+			require_db.logger.debug(`[Pricing] Cached pricing for ${provider}/${model}: input=$${pricing.inputCostPer1M}/1M, output=$${pricing.outputCostPer1M}/1M`);
+			return pricing;
 		} catch (error) {
-			require_db.logger.error(`[Pricing] Failed to fetch pricing data: ${error instanceof Error ? error.message : String(error)}`);
-			if (this.cache.size === 0) throw error;
+			require_db.logger.error(`[Pricing] Failed to fetch pricing for ${provider}/${model}: ${error instanceof Error ? error.message : String(error)}`);
+			const cacheKey = this.getCacheKey(provider, model);
+			const stale = this.cache.get(cacheKey);
+			if (stale) {
+				require_db.logger.debug(`[Pricing] Using stale cache for ${provider}/${model}`);
+				return stale.pricing;
+			}
+			return null;
 		}
 	}
 	/**
-	* Ensure cache is fresh, fetching if necessary
-	*/
-	async ensureFreshCache() {
-		if (!(Date.now() - this.lastFetch > this.cacheTTL) && this.cache.size > 0) return;
-		if (!this.fetchPromise) this.fetchPromise = this.fetchPricingData().finally(() => {
-			this.fetchPromise = null;
-		});
-		await this.fetchPromise;
-	}
-	/**
 	* Get pricing for a specific model
 	*/
 	async getModelPricing(provider, model) {
-		await this.ensureFreshCache();
 		const cacheKey = this.getCacheKey(provider, model);
-		const pricing = this.cache.get(cacheKey);
-		if (!pricing) {
-			require_db.logger.debug(`[Pricing] No pricing found for ${provider}/${model}, trying partial match`);
-			for (const [key, value] of this.cache.entries()) if (key.startsWith(`${provider.toLowerCase()}:`)) {
-				const modelPart = key.split(":")[1];
-				if (model.toLowerCase().includes(modelPart)) {
-					require_db.logger.debug(`[Pricing] Found partial match: ${key}`);
-					return value;
-				}
-			}
-			return null;
+		const cached = this.cache.get(cacheKey);
+		if (cached && Date.now() - cached.fetchedAt < this.cacheTTL) return cached.pricing;
+		let pending = this.pendingFetches.get(cacheKey);
+		if (!pending) {
+			pending = this.fetchModelPricing(provider, model).then((pricing) => {
+				this.cache.set(cacheKey, {
+					pricing,
+					fetchedAt: Date.now()
+				});
+				return pricing;
+			}).finally(() => {
+				this.pendingFetches.delete(cacheKey);
+			});
+			this.pendingFetches.set(cacheKey, pending);
 		}
-		return pricing;
+		return pending;
 	}
 	/**
-	* Force refresh the pricing cache
+	* Force refresh the pricing cache (clears all cached entries)
 	*/
 	async refreshCache() {
-		this.lastFetch = 0;
-		await this.ensureFreshCache();
+		this.cache.clear();
 	}
 	/**
-	* Check if the provider is ready
+	* Always ready — no bulk pre-fetch needed
 	*/
 	isReady() {
-		return this.ready;
+		return true;
 	}
 	/**
 	* Get the number of cached models (for debugging)
@@ -3882,7 +3944,7 @@ let defaultProvider = null;
 * Get the default pricing provider instance
 */
 function getDefaultPricingProvider() {
-	if (!defaultProvider) defaultProvider = new ModelsDevPricingProvider();
+	if (!defaultProvider) defaultProvider = new LLMOpsPricingProvider();
 	return defaultProvider;
 }
 
@@ -4267,14 +4329,15 @@ exports.LLMOPS_SPAN_NAME_HEADER = LLMOPS_SPAN_NAME_HEADER;
 exports.LLMOPS_TRACE_ID_HEADER = LLMOPS_TRACE_ID_HEADER;
 exports.LLMOPS_TRACE_NAME_HEADER = LLMOPS_TRACE_NAME_HEADER;
 exports.LLMOPS_USER_ID_HEADER = LLMOPS_USER_ID_HEADER;
+exports.LLMOpsPricingProvider = LLMOpsPricingProvider;
 exports.MS = MS;
 exports.ManifestBuilder = ManifestBuilder;
 exports.ManifestRouter = ManifestRouter;
 exports.ManifestService = ManifestService;
 exports.MemoryCacheBackend = MemoryCacheBackend;
-exports.ModelsDevPricingProvider = ModelsDevPricingProvider;
 exports.SCHEMA_METADATA = require_db.SCHEMA_METADATA;
 exports.SupportedProviders = SupportedProviders;
+exports.calculateCacheAwareCost = calculateCacheAwareCost;
 exports.calculateCost = calculateCost;
 exports.chatCompletionCreateParamsBaseSchema = chatCompletionCreateParamsBaseSchema;
 exports.configVariantsSchema = require_db.configVariantsSchema;

package/dist/index.d.cts

@@ -1,4 +1,4 @@
-import { $ as SCHEMA_METADATA, A as DatasetsTable, At as playgroundsSchema, B as LLMRequest, Bt as workspaceSettingsSchema, C as Dataset, Ct as environmentSecretsSchema, D as DatasetVersionRecord, Dt as playgroundColumnSchema, E as DatasetVersion, Et as llmRequestsSchema, F as GuardrailConfig, Ft as spansSchema, G as PlaygroundResultsTable, H as Playground, I as GuardrailConfigsTable, It as targetingRulesSchema, J as PlaygroundsTable, K as PlaygroundRun, L as GuardrailResult, Lt as tracesSchema, M as EnvironmentSecret, Mt as providerGuardrailOverridesSchema, N as EnvironmentSecretsTable, Nt as schemas, O as DatasetVersionRecordsTable, Ot as playgroundResultsSchema, P as EnvironmentsTable, Pt as spanEventsSchema, Q as ProviderGuardrailOverridesTable, R as GuardrailResults, Rt as variantVersionsSchema, S as Database, St as datasetsSchema, T as DatasetRecordsTable, Tt as guardrailConfigsSchema, U as PlaygroundColumn, V as LLMRequestsTable, W as PlaygroundResult, X as ProviderConfigsTable, Y as ProviderConfig, Z as ProviderGuardrailOverride, _ as validateTableData, _t as configVariantsSchema, a as createDatabaseFromConnection, at as TableName, b as ConfigVariantsTable, bt as datasetVersionRecordsSchema, c as executeWithSchema, ct as Trace, d as getMigrations, dt as Variant, et as Selectable, f as matchType, ft as VariantVersion, g as validatePartialTableData, gt as WorkspaceSettingsTable, h as parseTableData, ht as WorkspaceSettings, i as createDatabase, it as SpansTable, j as Environment, jt as providerConfigsSchema, k as DatasetVersionsTable, kt as playgroundRunsSchema, l as MigrationOptions, lt as TracesTable, m as parsePartialTableData, mt as VariantsTable, n as DatabaseOptions, nt as SpanEvent, o as detectDatabaseType, ot as TargetingRule, p as runAutoMigrations, pt as VariantVersionsTable, q as PlaygroundRunsTable, r as DatabaseType, rt as SpanEventsTable, s as createNeonDialect, st as TargetingRulesTable, t as DatabaseConnection, tt as Span, u as MigrationResult, ut as Updateable, v as Config, vt as configsSchema, w as DatasetRecord, wt as environmentsSchema, x as ConfigsTable, xt as datasetVersionsSchema, y as ConfigVariant, yt as datasetRecordsSchema, z as Insertable, zt as variantsSchema } from "./index-DdG7GtcE.cjs";
+import { $ as SCHEMA_METADATA, A as DatasetsTable, At as playgroundsSchema, B as LLMRequest, Bt as workspaceSettingsSchema, C as Dataset, Ct as environmentSecretsSchema, D as DatasetVersionRecord, Dt as playgroundColumnSchema, E as DatasetVersion, Et as llmRequestsSchema, F as GuardrailConfig, Ft as spansSchema, G as PlaygroundResultsTable, H as Playground, I as GuardrailConfigsTable, It as targetingRulesSchema, J as PlaygroundsTable, K as PlaygroundRun, L as GuardrailResult, Lt as tracesSchema, M as EnvironmentSecret, Mt as providerGuardrailOverridesSchema, N as EnvironmentSecretsTable, Nt as schemas, O as DatasetVersionRecordsTable, Ot as playgroundResultsSchema, P as EnvironmentsTable, Pt as spanEventsSchema, Q as ProviderGuardrailOverridesTable, R as GuardrailResults, Rt as variantVersionsSchema, S as Database, St as datasetsSchema, T as DatasetRecordsTable, Tt as guardrailConfigsSchema, U as PlaygroundColumn, V as LLMRequestsTable, W as PlaygroundResult, X as ProviderConfigsTable, Y as ProviderConfig, Z as ProviderGuardrailOverride, _ as validateTableData, _t as configVariantsSchema, a as createDatabaseFromConnection, at as TableName, b as ConfigVariantsTable, bt as datasetVersionRecordsSchema, c as executeWithSchema, ct as Trace, d as getMigrations, dt as Variant, et as Selectable, f as matchType, ft as VariantVersion, g as validatePartialTableData, gt as WorkspaceSettingsTable, h as parseTableData, ht as WorkspaceSettings, i as createDatabase, it as SpansTable, j as Environment, jt as providerConfigsSchema, k as DatasetVersionsTable, kt as playgroundRunsSchema, l as MigrationOptions, lt as TracesTable, m as parsePartialTableData, mt as VariantsTable, n as DatabaseOptions, nt as SpanEvent, o as detectDatabaseType, ot as TargetingRule, p as runAutoMigrations, pt as VariantVersionsTable, q as PlaygroundRunsTable, r as DatabaseType, rt as SpanEventsTable, s as createNeonDialect, st as TargetingRulesTable, t as DatabaseConnection, tt as Span, u as MigrationResult, ut as Updateable, v as Config, vt as configsSchema, w as DatasetRecord, wt as environmentsSchema, x as ConfigsTable, xt as datasetVersionsSchema, y as ConfigVariant, yt as datasetRecordsSchema, z as Insertable, zt as variantsSchema } from "./index-CyVFWq76.cjs";
 import * as kysely0 from "kysely";
 import { Kysely } from "kysely";
 import z$1, { z } from "zod";
@@ -1956,7 +1956,9 @@ declare const insertLLMRequestSchema: z$1.ZodObject<{
   completionTokens: z$1.ZodDefault<z$1.ZodNumber>;
   totalTokens: z$1.ZodDefault<z$1.ZodNumber>;
   cachedTokens: z$1.ZodDefault<z$1.ZodNumber>;
+  cacheCreationTokens: z$1.ZodDefault<z$1.ZodNumber>;
   cost: z$1.ZodDefault<z$1.ZodNumber>;
+  cacheSavings: z$1.ZodDefault<z$1.ZodNumber>;
   inputCost: z$1.ZodDefault<z$1.ZodNumber>;
   outputCost: z$1.ZodDefault<z$1.ZodNumber>;
   endpoint: z$1.ZodString;
@@ -2069,6 +2071,8 @@ declare const createLLMRequestsDataLayer: (db: Kysely<Database>) => {
     requestId: string;
     model: string;
     cachedTokens: number;
+    cacheCreationTokens: number;
+    cacheSavings: number;
     inputCost: number;
     outputCost: number;
     endpoint: string;
@@ -2114,6 +2118,8 @@ declare const createLLMRequestsDataLayer: (db: Kysely<Database>) => {
       requestId: string;
       model: string;
       cachedTokens: number;
+      cacheCreationTokens: number;
+      cacheSavings: number;
       inputCost: number;
       outputCost: number;
       endpoint: string;
@@ -2161,6 +2167,8 @@ declare const createLLMRequestsDataLayer: (db: Kysely<Database>) => {
     requestId: string;
     model: string;
     cachedTokens: number;
+    cacheCreationTokens: number;
+    cacheSavings: number;
     inputCost: number;
     outputCost: number;
     endpoint: string;
@@ -2194,6 +2202,8 @@ declare const createLLMRequestsDataLayer: (db: Kysely<Database>) => {
     totalPromptTokens: number;
     totalCompletionTokens: number;
     totalTokens: number;
+    totalCachedTokens: number;
+    totalCacheSavings: number;
     requestCount: number;
   } | undefined>;
   /**
@@ -3787,8 +3797,10 @@ interface UsageData {
   completionTokens: number;
   /** Total tokens (prompt + completion) */
   totalTokens?: number;
-  /** Number of cached tokens (optional) */
+  /** Number of cache read tokens (OpenAI cached_tokens / Anthropic cache_read_input_tokens) */
   cachedTokens?: number;
+  /** Number of cache creation tokens (Anthropic cache_creation_input_tokens) */
+  cacheCreationTokens?: number;
   /** Number of reasoning tokens (optional, for models like o1) */
   reasoningTokens?: number;
 }
@@ -3804,6 +3816,8 @@ interface CostResult {
   inputCost: number;
   /** Output/completion cost in micro-dollars */
   outputCost: number;
+  /** Cost saved by cache hits in micro-dollars (negative means cache write premium exceeded savings) */
+  cacheSavings: number;
 }
 /**
  * Provider for fetching model pricing data
@@ -3850,6 +3864,19 @@ interface PricingProvider {
  * ```
  */
 declare function calculateCost(usage: UsageData, pricing: ModelPricing): CostResult;
+/**
+ * Calculate cache-aware cost of an LLM request in micro-dollars.
+ *
+ * Splits input tokens into uncached, cache-read, and cache-creation buckets,
+ * each priced at different rates. Falls back to provider-specific multipliers
+ * when models.dev doesn't provide cache pricing.
+ *
+ * @param usage - Token usage data (with cachedTokens and cacheCreationTokens)
+ * @param pricing - Model pricing (may include cacheReadCostPer1M / cacheWriteCostPer1M)
+ * @param provider - Provider name for fallback rate selection
+ * @returns Cost breakdown in micro-dollars
+ */
+declare function calculateCacheAwareCost(usage: UsageData, pricing: ModelPricing, provider?: string): CostResult;
 /**
  * Convert micro-dollars to dollars
  *
@@ -3893,47 +3920,38 @@ declare function formatCost(microDollars: number, decimals?: number): string;
 //#endregion
 //#region src/pricing/provider.d.ts
 /**
- * Pricing provider that fetches data from models.dev API
+ * Pricing provider that fetches per-model data from the LLMOps Models API.
  *
  * Features:
- * - Caches pricing data with configurable TTL (default 5 minutes)
- * - Supports fallback to local cache on fetch failure
- * - Thread-safe cache refresh
+ * - Per-model in-memory cache with configurable TTL (default 5 minutes)
+ * - Deduplicates concurrent fetches for the same model
+ * - Caches null results (404s) to avoid repeated lookups
+ * - Falls back to stale cache on fetch errors
  */
-declare class ModelsDevPricingProvider implements PricingProvider {
+declare class LLMOpsPricingProvider implements PricingProvider {
   private cache;
-  private lastFetch;
+  private pendingFetches;
   private cacheTTL;
-  private fetchPromise;
-  private ready;
-  /**
-   * Create a new ModelsDevPricingProvider
-   *
-   * @param cacheTTL - Cache TTL in milliseconds (default: 5 minutes)
-   */
-  constructor(cacheTTL?: number);
-  /**
-   * Generate a cache key for a provider/model combination
-   */
+  private baseUrl;
+  constructor(options?: {
+    cacheTTL?: number;
+    baseUrl?: string;
+  });
   private getCacheKey;
   /**
-   * Fetch pricing data from models.dev API
+   * Fetch pricing for a single model from the API
    */
-  private fetchPricingData;
-  /**
-   * Ensure cache is fresh, fetching if necessary
-   */
-  private ensureFreshCache;
+  private fetchModelPricing;
   /**
    * Get pricing for a specific model
    */
   getModelPricing(provider: string, model: string): Promise<ModelPricing | null>;
   /**
-   * Force refresh the pricing cache
+   * Force refresh the pricing cache (clears all cached entries)
    */
   refreshCache(): Promise<void>;
   /**
-   * Check if the provider is ready
+   * Always ready — no bulk pre-fetch needed
    */
   isReady(): boolean;
   /**
@@ -3944,7 +3962,7 @@ declare class ModelsDevPricingProvider implements PricingProvider {
 /**
  * Get the default pricing provider instance
  */
-declare function getDefaultPricingProvider(): ModelsDevPricingProvider;
+declare function getDefaultPricingProvider(): LLMOpsPricingProvider;
 //#endregion
 //#region src/auth/get-auth-client-options.d.ts
 interface AuthClientDatabaseConfig {
@@ -4174,4 +4192,4 @@ declare class ManifestRouter {
   routeWithWeights(configIdOrSlug: string, environmentId: string, context?: RoutingContext): RoutingResult | null;
 }
 //#endregion
-export { type AnthropicProviderConfig, type AnyProviderConfig, AuthClientDatabaseConfig, AuthClientOptions, type AzureAIProviderConfig, type AzureOpenAIProviderConfig, BaseCacheConfig, type BaseProviderConfig, type BedrockProviderConfig, COST_SUMMARY_GROUP_BY, CacheBackend, CacheBackendType, CacheConfig, CacheEntry, CacheOptions, CacheService, CacheStats, ChatCompletionCreateParamsBase, Config, ConfigVariant, ConfigVariantsDataLayer, ConfigVariantsTable, ConfigsDataLayer, ConfigsTable, type CortexProviderConfig, CostResult, CostSummaryGroupBy, DEFAULT_PROVIDER_ENV_VARS, DataLayer, Database, DatabaseConnection, DatabaseOptions, DatabaseType, Dataset, DatasetRecord, DatasetRecordsTable, DatasetVersion, DatasetVersionRecord, DatasetVersionRecordsTable, DatasetVersionsTable, DatasetsDataLayer, DatasetsTable, Environment, EnvironmentSecret, EnvironmentSecretsDataLayer, EnvironmentSecretsTable, EnvironmentsDataLayer, EnvironmentsTable, FileCacheBackend, FileCacheConfig, type FireworksAIProviderConfig, GatewayManifest, type GoogleProviderConfig, GuardrailConfig, GuardrailConfigsDataLayer, GuardrailConfigsTable, GuardrailResult, GuardrailResults, type HuggingFaceProviderConfig, type InlineProviderConfig, type InlineProvidersConfig, Insertable, LLMOPS_INTERNAL_HEADER, LLMOPS_REQUEST_ID_HEADER, LLMOPS_SESSION_ID_HEADER, LLMOPS_SPAN_ID_HEADER, LLMOPS_SPAN_NAME_HEADER, LLMOPS_TRACE_ID_HEADER, LLMOPS_TRACE_NAME_HEADER, LLMOPS_USER_ID_HEADER, LLMOpsClient, LLMOpsConfig, type LLMOpsConfigInput, LLMRequest, LLMRequestInsert, LLMRequestsDataLayer, LLMRequestsTable, MS, ManifestBuilder, ManifestConfig, ManifestEnvironment, ManifestGuardrail, ManifestProviderGuardrailOverride, ManifestRouter, ManifestService, ManifestTargetingRule, ManifestVariantVersion, MemoryCacheBackend, MemoryCacheConfig, MigrationOptions, MigrationResult, type MistralAIProviderConfig, ModelPricing, ModelsDevPricingProvider, type OpenAIProviderConfig, type OracleProviderConfig, Playground, PlaygroundColumn, PlaygroundResult, PlaygroundResultsDataLayer, PlaygroundResultsTable, PlaygroundRun, PlaygroundRunsDataLayer, PlaygroundRunsTable, PlaygroundsDataLayer, PlaygroundsTable, Prettify, PricingProvider, ProviderConfig, type ProviderConfigMap, ProviderConfigsDataLayer, ProviderConfigsTable, ProviderGuardrailOverride, ProviderGuardrailOverridesDataLayer, ProviderGuardrailOverridesTable, type ProvidersConfig, RoutingContext, RoutingResult, SCHEMA_METADATA, type SagemakerProviderConfig, Selectable, Span, SpanEvent, SpanEventInsert, SpanEventsTable, SpanInsert, SpansTable, type StabilityAIProviderConfig, SupportedProviders, TableName, TargetingRule, TargetingRulesDataLayer, TargetingRulesTable, Trace, TraceUpsert, TracesDataLayer, TracesTable, Updateable, UsageData, type ValidatedLLMOpsConfig, Variant, VariantJsonData, VariantVersion, VariantVersionsDataLayer, VariantVersionsTable, VariantsDataLayer, VariantsTable, type VertexAIProviderConfig, type WorkersAIProviderConfig, WorkspaceSettings, WorkspaceSettingsDataLayer, WorkspaceSettingsTable, calculateCost, chatCompletionCreateParamsBaseSchema, configVariantsSchema, configsSchema, createConfigDataLayer, createConfigVariantDataLayer, createDataLayer, createDatabase, createDatabaseFromConnection, createDatasetsDataLayer, createEnvironmentDataLayer, createEnvironmentSecretDataLayer, createGuardrailConfigsDataLayer, createLLMRequestsDataLayer, createNeonDialect, createPlaygroundDataLayer, createPlaygroundResultsDataLayer, createPlaygroundRunsDataLayer, createProviderConfigsDataLayer, createProviderGuardrailOverridesDataLayer, createTargetingRulesDataLayer, createTracesDataLayer, createVariantDataLayer, createVariantVersionsDataLayer, createWorkspaceSettingsDataLayer, datasetRecordsSchema, datasetVersionRecordsSchema, datasetVersionsSchema, datasetsSchema, detectDatabaseType, dollarsToMicroDollars, environmentSecretsSchema, environmentsSchema, executeWithSchema, formatCost, gateway, generateId, getAuthClientOptions, getDefaultPricingProvider, getDefaultProviders, getMigrations, guardrailConfigsSchema, llmRequestsSchema, llmopsConfigSchema, logger, matchType, mergeWithDefaultProviders, microDollarsToDollars, parsePartialTableData, parseTableData, playgroundColumnSchema, playgroundResultsSchema, playgroundRunsSchema, playgroundsSchema, providerConfigsSchema, providerGuardrailOverridesSchema, runAutoMigrations, schemas, spanEventsSchema, spansSchema, targetingRulesSchema, tracesSchema, validateLLMOpsConfig, validatePartialTableData, validateTableData, variantJsonDataSchema, variantVersionsSchema, variantsSchema, workspaceSettingsSchema };
+export { type AnthropicProviderConfig, type AnyProviderConfig, AuthClientDatabaseConfig, AuthClientOptions, type AzureAIProviderConfig, type AzureOpenAIProviderConfig, BaseCacheConfig, type BaseProviderConfig, type BedrockProviderConfig, COST_SUMMARY_GROUP_BY, CacheBackend, CacheBackendType, CacheConfig, CacheEntry, CacheOptions, CacheService, CacheStats, ChatCompletionCreateParamsBase, Config, ConfigVariant, ConfigVariantsDataLayer, ConfigVariantsTable, ConfigsDataLayer, ConfigsTable, type CortexProviderConfig, CostResult, CostSummaryGroupBy, DEFAULT_PROVIDER_ENV_VARS, DataLayer, Database, DatabaseConnection, DatabaseOptions, DatabaseType, Dataset, DatasetRecord, DatasetRecordsTable, DatasetVersion, DatasetVersionRecord, DatasetVersionRecordsTable, DatasetVersionsTable, DatasetsDataLayer, DatasetsTable, Environment, EnvironmentSecret, EnvironmentSecretsDataLayer, EnvironmentSecretsTable, EnvironmentsDataLayer, EnvironmentsTable, FileCacheBackend, FileCacheConfig, type FireworksAIProviderConfig, GatewayManifest, type GoogleProviderConfig, GuardrailConfig, GuardrailConfigsDataLayer, GuardrailConfigsTable, GuardrailResult, GuardrailResults, type HuggingFaceProviderConfig, type InlineProviderConfig, type InlineProvidersConfig, Insertable, LLMOPS_INTERNAL_HEADER, LLMOPS_REQUEST_ID_HEADER, LLMOPS_SESSION_ID_HEADER, LLMOPS_SPAN_ID_HEADER, LLMOPS_SPAN_NAME_HEADER, LLMOPS_TRACE_ID_HEADER, LLMOPS_TRACE_NAME_HEADER, LLMOPS_USER_ID_HEADER, LLMOpsClient, LLMOpsConfig, type LLMOpsConfigInput, LLMOpsPricingProvider, LLMRequest, LLMRequestInsert, LLMRequestsDataLayer, LLMRequestsTable, MS, ManifestBuilder, ManifestConfig, ManifestEnvironment, ManifestGuardrail, ManifestProviderGuardrailOverride, ManifestRouter, ManifestService, ManifestTargetingRule, ManifestVariantVersion, MemoryCacheBackend, MemoryCacheConfig, MigrationOptions, MigrationResult, type MistralAIProviderConfig, ModelPricing, type OpenAIProviderConfig, type OracleProviderConfig, Playground, PlaygroundColumn, PlaygroundResult, PlaygroundResultsDataLayer, PlaygroundResultsTable, PlaygroundRun, PlaygroundRunsDataLayer, PlaygroundRunsTable, PlaygroundsDataLayer, PlaygroundsTable, Prettify, PricingProvider, ProviderConfig, type ProviderConfigMap, ProviderConfigsDataLayer, ProviderConfigsTable, ProviderGuardrailOverride, ProviderGuardrailOverridesDataLayer, ProviderGuardrailOverridesTable, type ProvidersConfig, RoutingContext, RoutingResult, SCHEMA_METADATA, type SagemakerProviderConfig, Selectable, Span, SpanEvent, SpanEventInsert, SpanEventsTable, SpanInsert, SpansTable, type StabilityAIProviderConfig, SupportedProviders, TableName, TargetingRule, TargetingRulesDataLayer, TargetingRulesTable, Trace, TraceUpsert, TracesDataLayer, TracesTable, Updateable, UsageData, type ValidatedLLMOpsConfig, Variant, VariantJsonData, VariantVersion, VariantVersionsDataLayer, VariantVersionsTable, VariantsDataLayer, VariantsTable, type VertexAIProviderConfig, type WorkersAIProviderConfig, WorkspaceSettings, WorkspaceSettingsDataLayer, WorkspaceSettingsTable, calculateCacheAwareCost, calculateCost, chatCompletionCreateParamsBaseSchema, configVariantsSchema, configsSchema, createConfigDataLayer, createConfigVariantDataLayer, createDataLayer, createDatabase, createDatabaseFromConnection, createDatasetsDataLayer, createEnvironmentDataLayer, createEnvironmentSecretDataLayer, createGuardrailConfigsDataLayer, createLLMRequestsDataLayer, createNeonDialect, createPlaygroundDataLayer, createPlaygroundResultsDataLayer, createPlaygroundRunsDataLayer, createProviderConfigsDataLayer, createProviderGuardrailOverridesDataLayer, createTargetingRulesDataLayer, createTracesDataLayer, createVariantDataLayer, createVariantVersionsDataLayer, createWorkspaceSettingsDataLayer, datasetRecordsSchema, datasetVersionRecordsSchema, datasetVersionsSchema, datasetsSchema, detectDatabaseType, dollarsToMicroDollars, environmentSecretsSchema, environmentsSchema, executeWithSchema, formatCost, gateway, generateId, getAuthClientOptions, getDefaultPricingProvider, getDefaultProviders, getMigrations, guardrailConfigsSchema, llmRequestsSchema, llmopsConfigSchema, logger, matchType, mergeWithDefaultProviders, microDollarsToDollars, parsePartialTableData, parseTableData, playgroundColumnSchema, playgroundResultsSchema, playgroundRunsSchema, playgroundsSchema, providerConfigsSchema, providerGuardrailOverridesSchema, runAutoMigrations, schemas, spanEventsSchema, spansSchema, targetingRulesSchema, tracesSchema, validateLLMOpsConfig, validatePartialTableData, validateTableData, variantJsonDataSchema, variantVersionsSchema, variantsSchema, workspaceSettingsSchema };