@mastra/observability 1.17.1-alpha.0 → 1.17.1-alpha.1

package/CHANGELOG.md

@@ -1,5 +1,14 @@
 # @mastra/observability
 
+## 1.17.1-alpha.1
+
+### Patch Changes
+
+- Fixed Anthropic cache-write cost estimates by applying TTL-specific rates without double counting aggregate tokens. ([#21563](https://github.com/mastra-ai/mastra/pull/21563))
+
+- Updated dependencies [[`d7e6745`](https://github.com/mastra-ai/mastra/commit/d7e67456954863c55440ea9c49bc6ceb9949972d), [`9acb50f`](https://github.com/mastra-ai/mastra/commit/9acb50f71cec9c362f06820033f90ae6b1f8282f), [`46e9e3f`](https://github.com/mastra-ai/mastra/commit/46e9e3f73babe1bc70080a596cf2ac0b9da48519), [`3f9a190`](https://github.com/mastra-ai/mastra/commit/3f9a19057c027155867b9317294ee4ca7bd0581a), [`e8808e3`](https://github.com/mastra-ai/mastra/commit/e8808e3d8eb585a2565be53e56a7e0e1477352a4), [`d4be8c1`](https://github.com/mastra-ai/mastra/commit/d4be8c1739d22d621e3f78790e1dd5eb5ecc3589), [`a5d2eb1`](https://github.com/mastra-ai/mastra/commit/a5d2eb10347eade1ae2816d88f466c25186c54a5), [`e81744c`](https://github.com/mastra-ai/mastra/commit/e81744cd13c46619c142dc521dc0baac47607a84)]:
+  - @mastra/core@1.60.0-alpha.4
+
 ## 1.17.1-alpha.0
 
 ### Patch Changes

package/README.md

@@ -56,9 +56,11 @@ Metrics are automatically extracted from span lifecycle events by `AutoExtracted
 - `mastra_workflow_duration_ms`
 - `mastra_model_duration_ms`
 - `mastra_model_total_input_tokens` / `mastra_model_total_output_tokens`
-- `mastra_model_input_text_tokens` / `mastra_model_input_cache_read_tokens` / `mastra_model_input_cache_write_tokens` / `mastra_model_input_audio_tokens` / `mastra_model_input_image_tokens`
+- `mastra_model_input_text_tokens` / `mastra_model_input_cache_read_tokens` / `mastra_model_input_cache_write_tokens` / `mastra_model_input_cache_write_5m_tokens` / `mastra_model_input_cache_write_1h_tokens` / `mastra_model_input_audio_tokens` / `mastra_model_input_image_tokens`
 - `mastra_model_output_text_tokens` / `mastra_model_output_reasoning_tokens` / `mastra_model_output_audio_tokens` / `mastra_model_output_image_tokens`
 
+For Anthropic models, the aggregate cache-write metric remains available while the 5-minute and 1-hour metrics preserve the provider's TTL-specific token counts and pricing.
+
 Auto-extracted metrics carry labels: `entity_type`, `entity_name`, `status`, plus `model` and `provider` on model generation spans.
 
 ### Structured logging

package/dist/index.cjs

@@ -1102,6 +1102,8 @@ const MINIFIED_METER_TO_CANONICAL = {
 	ot: "output_tokens",
 	icrt: "input_cache_read_tokens",
 	icwt: "input_cache_write_tokens",
+	icwt5m: "input_cache_write_5m_tokens",
+	icwt1h: "input_cache_write_1h_tokens",
 	iat: "input_audio_tokens",
 	oat: "output_audio_tokens",
 	ort: "output_reasoning_tokens"
@@ -1158,6 +1160,15 @@ function parsePricingModelText(content) {
 	}
 	return pricingModels;
 }
+function expandRates(row, tier) {
+	const rates = Object.fromEntries(Object.entries(tier.r).map(([meter, value]) => [MINIFIED_METER_TO_CANONICAL[meter], value.c]));
+	const cacheWriteRate = rates.input_cache_write_tokens;
+	if (row.m.includes("claude") && typeof cacheWriteRate === "number") {
+		rates.input_cache_write_5m_tokens ??= cacheWriteRate;
+		rates.input_cache_write_1h_tokens ??= cacheWriteRate * 1.6;
+	}
+	return rates;
+}
 function expandPricingModelRow(row) {
 	return new PricingModel({
 		id: row.i,
@@ -1172,7 +1183,7 @@ function expandPricingModelRow(row) {
 				op: condition.op,
 				value: condition.value
 			})),
-			rates: Object.fromEntries(Object.entries(tier.r).map(([meter, value]) => [MINIFIED_METER_TO_CANONICAL[meter], value.c]))
+			rates: expandRates(row, tier)
 		}))
 	});
 }
@@ -1303,6 +1314,8 @@ const PricingMeter = {
 	INPUT_AUDIO_TOKENS: "input_audio_tokens",
 	INPUT_CACHE_READ_TOKENS: "input_cache_read_tokens",
 	INPUT_CACHE_WRITE_TOKENS: "input_cache_write_tokens",
+	INPUT_CACHE_WRITE_5M_TOKENS: "input_cache_write_5m_tokens",
+	INPUT_CACHE_WRITE_1H_TOKENS: "input_cache_write_1h_tokens",
 	INPUT_IMAGE_TOKENS: "input_image_tokens",
 	OUTPUT_TOKENS: "output_tokens",
 	OUTPUT_AUDIO_TOKENS: "output_audio_tokens",
@@ -1315,6 +1328,8 @@ const TokenMetrics = {
 	INPUT_TEXT: "mastra_model_input_text_tokens",
 	INPUT_CACHE_READ: "mastra_model_input_cache_read_tokens",
 	INPUT_CACHE_WRITE: "mastra_model_input_cache_write_tokens",
+	INPUT_CACHE_WRITE_5M: "mastra_model_input_cache_write_5m_tokens",
+	INPUT_CACHE_WRITE_1H: "mastra_model_input_cache_write_1h_tokens",
 	INPUT_AUDIO: "mastra_model_input_audio_tokens",
 	INPUT_IMAGE: "mastra_model_input_image_tokens",
 	OUTPUT_TEXT: "mastra_model_output_text_tokens",
@@ -1344,6 +1359,8 @@ function getTokenMetricSamples(usage) {
 		pushIfPositive(TokenMetrics.INPUT_TEXT, usage.inputDetails.text);
 		pushIfPositive(TokenMetrics.INPUT_CACHE_READ, usage.inputDetails.cacheRead);
 		pushIfPositive(TokenMetrics.INPUT_CACHE_WRITE, usage.inputDetails.cacheWrite);
+		pushIfPositive(TokenMetrics.INPUT_CACHE_WRITE_5M, usage.inputDetails.cacheWrite5m);
+		pushIfPositive(TokenMetrics.INPUT_CACHE_WRITE_1H, usage.inputDetails.cacheWrite1h);
 		pushIfPositive(TokenMetrics.INPUT_AUDIO, usage.inputDetails.audio);
 		pushIfPositive(TokenMetrics.INPUT_IMAGE, usage.inputDetails.image);
 	}
@@ -1410,15 +1427,40 @@ function estimateCosts(args, pricingRegistry = PricingRegistry.getGlobal()) {
 		results.set(TokenMetrics.INPUT_CACHE_READ, result.costContext);
 		inputDetailResults.push(result);
 	}
-	if (usage.inputDetails?.cacheWrite) {
+	const cacheWriteDetailResults = [];
+	const cacheWrite5m = usage.inputDetails?.cacheWrite5m ?? 0;
+	const cacheWrite1h = usage.inputDetails?.cacheWrite1h ?? 0;
+	if (cacheWrite5m > 0) {
+		const result = estimateCostForMeter({
+			meter: PricingMeter.INPUT_CACHE_WRITE_5M_TOKENS,
+			tokenCount: cacheWrite5m,
+			...estimateFields
+		});
+		results.set(TokenMetrics.INPUT_CACHE_WRITE_5M, result.costContext);
+		cacheWriteDetailResults.push(result);
+		inputDetailResults.push(result);
+	}
+	if (cacheWrite1h > 0) {
+		const result = estimateCostForMeter({
+			meter: PricingMeter.INPUT_CACHE_WRITE_1H_TOKENS,
+			tokenCount: cacheWrite1h,
+			...estimateFields
+		});
+		results.set(TokenMetrics.INPUT_CACHE_WRITE_1H, result.costContext);
+		cacheWriteDetailResults.push(result);
+		inputDetailResults.push(result);
+	}
+	const unclassifiedCacheWrite = Math.max(0, (usage.inputDetails?.cacheWrite ?? 0) - cacheWrite5m - cacheWrite1h);
+	if (unclassifiedCacheWrite > 0) {
 		const result = estimateCostForMeter({
 			meter: PricingMeter.INPUT_CACHE_WRITE_TOKENS,
-			tokenCount: usage.inputDetails.cacheWrite,
+			tokenCount: unclassifiedCacheWrite,
 			...estimateFields
 		});
-		results.set(TokenMetrics.INPUT_CACHE_WRITE, result.costContext);
+		cacheWriteDetailResults.push(result);
 		inputDetailResults.push(result);
 	}
+	if (cacheWriteDetailResults.length > 0) setCombinedCostContext(results, TokenMetrics.INPUT_CACHE_WRITE, cacheWriteDetailResults, pricingModel, costMetadata);
 	if (usage.inputDetails?.image) {
 		const result = estimateCostForMeter({
 			meter: PricingMeter.INPUT_IMAGE_TOKENS,
@@ -1495,6 +1537,20 @@ function estimateCosts(args, pricingRegistry = PricingRegistry.getGlobal()) {
 function applyErrorContextForUsage(results, usage, errorContext) {
 	for (const sample of getTokenMetricSamples(usage)) results.set(sample.name, errorContext);
 }
+function setCombinedCostContext(results, metric, detailResults, pricingModel, costMetadata) {
+	const estimatedCosts = detailResults.map((result) => result.costContext.estimatedCost).filter((value) => typeof value === "number");
+	const hasFailedCost = detailResults.some((result) => !result.success);
+	results.set(metric, {
+		provider: pricingModel.provider,
+		model: pricingModel.model,
+		...estimatedCosts.length > 0 && { estimatedCost: estimatedCosts.reduce((sum, value) => sum + value, 0) },
+		...estimatedCosts.length > 0 && { costUnit: pricingModel.currency },
+		costMetadata: hasFailedCost ? {
+			...costMetadata,
+			error: "partial_cost"
+		} : { ...costMetadata }
+	});
+}
 function setAggregateCostContext(args) {
 	const { results, totalMetric, fallbackMeter, totalTokenCount, detailResults, pricingModel, pricingTier, costMetadata } = args;
 	if (totalTokenCount == null) return;
@@ -1735,14 +1791,24 @@ function extractUsageMetrics(usage, providerMetadata) {
 	if (isDefined(aiSdkDetails?.cacheReadTokens)) inputDetails.cacheRead = aiSdkDetails.cacheReadTokens;
 	if (isDefined(aiSdkDetails?.cacheWriteTokens)) inputDetails.cacheWrite = aiSdkDetails.cacheWriteTokens;
 	if (!isDefined(inputDetails.cacheRead) && isDefined(usage.cachedInputTokens)) inputDetails.cacheRead = usage.cachedInputTokens;
+	if (isDefined(usage.cacheCreationInputTokens5m)) inputDetails.cacheWrite5m = usage.cacheCreationInputTokens5m;
+	if (isDefined(usage.cacheCreationInputTokens1h)) inputDetails.cacheWrite1h = usage.cacheCreationInputTokens1h;
 	if (!isDefined(inputDetails.cacheWrite) && isDefined(usage.cacheCreationInputTokens)) inputDetails.cacheWrite = usage.cacheCreationInputTokens;
+	if (!isDefined(inputDetails.cacheWrite) && (isDefined(inputDetails.cacheWrite5m) || isDefined(inputDetails.cacheWrite1h))) inputDetails.cacheWrite = (inputDetails.cacheWrite5m ?? 0) + (inputDetails.cacheWrite1h ?? 0);
 	if (isDefined(usage.reasoningTokens)) outputDetails.reasoning = usage.reasoningTokens;
 	const anthropic = providerMetadata?.anthropic;
 	if (anthropic) {
 		const rawV3InputUsage = isV3RawUsage(usage.raw) ? usage.raw.inputTokens : void 0;
 		const hasV3CachedTotals = rawV3InputUsage?.total !== void 0 && (rawV3InputUsage.cacheRead !== void 0 || rawV3InputUsage.cacheWrite !== void 0);
 		if (!isDefined(inputDetails.cacheRead) && isDefined(anthropic.cacheReadInputTokens)) inputDetails.cacheRead = anthropic.cacheReadInputTokens;
-		if (!isDefined(inputDetails.cacheWrite) && isDefined(anthropic.cacheCreationInputTokens)) inputDetails.cacheWrite = anthropic.cacheCreationInputTokens;
+		const cacheWrite5m = anthropic.cacheCreation?.ephemeral_5m_input_tokens ?? anthropic.cacheCreation?.ephemeral5mInputTokens;
+		const cacheWrite1h = anthropic.cacheCreation?.ephemeral_1h_input_tokens ?? anthropic.cacheCreation?.ephemeral1hInputTokens;
+		if (!isDefined(inputDetails.cacheWrite5m) && isDefined(cacheWrite5m)) inputDetails.cacheWrite5m = cacheWrite5m;
+		if (!isDefined(inputDetails.cacheWrite1h) && isDefined(cacheWrite1h)) inputDetails.cacheWrite1h = cacheWrite1h;
+		if (!isDefined(inputDetails.cacheWrite)) {
+			if (isDefined(anthropic.cacheCreationInputTokens)) inputDetails.cacheWrite = anthropic.cacheCreationInputTokens;
+			else if (isDefined(cacheWrite5m) || isDefined(cacheWrite1h)) inputDetails.cacheWrite = (cacheWrite5m ?? 0) + (cacheWrite1h ?? 0);
+		}
 		if (!(hasV3CachedTotals || isDefined(usage.cachedInputTokens) && usage.cachedInputTokens > 0 || isDefined(usage.cacheCreationInputTokens) && usage.cacheCreationInputTokens > 0) && (isDefined(inputDetails.cacheRead) || isDefined(inputDetails.cacheWrite))) inputTokens = (usage.inputTokens ?? 0) + (inputDetails.cacheRead ?? 0) + (inputDetails.cacheWrite ?? 0);
 	}
 	const google = providerMetadata?.google;
@@ -1783,6 +1849,8 @@ function mergeInputDetails(a, b) {
 		text: addOptional(a.text, b.text),
 		cacheRead: addOptional(a.cacheRead, b.cacheRead),
 		cacheWrite: addOptional(a.cacheWrite, b.cacheWrite),
+		cacheWrite5m: addOptional(a.cacheWrite5m, b.cacheWrite5m),
+		cacheWrite1h: addOptional(a.cacheWrite1h, b.cacheWrite1h),
 		audio: addOptional(a.audio, b.audio),
 		image: addOptional(a.image, b.image)
 	};