fix(core): Fix and add missing cache attributes in Vercel AI (#17982)

With the relay now handling cache token attributes (instead of scrubbing
them), some Anthropic related token attributes were still missing. This
PR adds the missing cache attributes and corrects the types in the
Anthropic provider metadata used for extracting token data.

Fixes: #17890

Author: RulaKhaled

packages/core/src/utils/ai/gen-ai-attributes.ts

@@ -129,6 +129,16 @@ export const GEN_AI_RESPONSE_STREAMING_ATTRIBUTE = 'gen_ai.response.streaming';
  */
 export const GEN_AI_RESPONSE_TOOL_CALLS_ATTRIBUTE = 'gen_ai.response.tool_calls';
 
+/**
+ * The number of cache write input tokens used
+ */
+export const GEN_AI_USAGE_INPUT_TOKENS_CACHE_WRITE_ATTRIBUTE = 'gen_ai.usage.input_tokens.cache_write';
+
+/**
+ * The number of cached input tokens that were used
+ */
+export const GEN_AI_USAGE_INPUT_TOKENS_CACHED_ATTRIBUTE = 'gen_ai.usage.input_tokens.cached';
+
 // =============================================================================
 // OPENAI-SPECIFIC ATTRIBUTES
 // =============================================================================

packages/core/src/utils/vercel-ai/index.ts

@@ -2,6 +2,10 @@ import type { Client } from '../../client';
 import { SEMANTIC_ATTRIBUTE_SENTRY_OP, SEMANTIC_ATTRIBUTE_SENTRY_ORIGIN } from '../../semanticAttributes';
 import type { Event } from '../../types-hoist/event';
 import type { Span, SpanAttributes, SpanAttributeValue, SpanJSON, SpanOrigin } from '../../types-hoist/span';
+import {
+  GEN_AI_USAGE_INPUT_TOKENS_CACHE_WRITE_ATTRIBUTE,
+  GEN_AI_USAGE_INPUT_TOKENS_CACHED_ATTRIBUTE,
+} from '../ai/gen-ai-attributes';
 import { spanToJSON } from '../spanUtils';
 import { toolCallSpanMap } from './constants';
 import type { TokenSummary } from './types';
@@ -23,6 +27,7 @@ import {
   AI_TOOL_CALL_ID_ATTRIBUTE,
   AI_TOOL_CALL_NAME_ATTRIBUTE,
   AI_TOOL_CALL_RESULT_ATTRIBUTE,
+  AI_USAGE_CACHED_INPUT_TOKENS_ATTRIBUTE,
   AI_USAGE_COMPLETION_TOKENS_ATTRIBUTE,
   AI_USAGE_PROMPT_TOKENS_ATTRIBUTE,
   GEN_AI_RESPONSE_MODEL_ATTRIBUTE,
@@ -107,6 +112,7 @@ function processEndedVercelAiSpan(span: SpanJSON): void {
 
   renameAttributeKey(attributes, AI_USAGE_COMPLETION_TOKENS_ATTRIBUTE, GEN_AI_USAGE_OUTPUT_TOKENS_ATTRIBUTE);
   renameAttributeKey(attributes, AI_USAGE_PROMPT_TOKENS_ATTRIBUTE, GEN_AI_USAGE_INPUT_TOKENS_ATTRIBUTE);
+  renameAttributeKey(attributes, AI_USAGE_CACHED_INPUT_TOKENS_ATTRIBUTE, GEN_AI_USAGE_INPUT_TOKENS_CACHED_ATTRIBUTE);
 
   if (
     typeof attributes[GEN_AI_USAGE_OUTPUT_TOKENS_ATTRIBUTE] === 'number' &&
@@ -287,7 +293,7 @@ function addProviderMetadataToAttributes(attributes: SpanAttributes): void {
       if (providerMetadataObject.openai) {
         setAttributeIfDefined(
           attributes,
-          'gen_ai.usage.input_tokens.cached',
+          GEN_AI_USAGE_INPUT_TOKENS_CACHED_ATTRIBUTE,
           providerMetadataObject.openai.cachedPromptTokens,
         );
         setAttributeIfDefined(
@@ -309,35 +315,34 @@ function addProviderMetadataToAttributes(attributes: SpanAttributes): void {
       }
 
       if (providerMetadataObject.anthropic) {
-        setAttributeIfDefined(
-          attributes,
-          'gen_ai.usage.input_tokens.cached',
-          providerMetadataObject.anthropic.cacheReadInputTokens,
-        );
-        setAttributeIfDefined(
-          attributes,
-          'gen_ai.usage.input_tokens.cache_write',
-          providerMetadataObject.anthropic.cacheCreationInputTokens,
-        );
+        const cachedInputTokens =
+          providerMetadataObject.anthropic.usage?.cache_read_input_tokens ??
+          providerMetadataObject.anthropic.cacheReadInputTokens;
+        setAttributeIfDefined(attributes, GEN_AI_USAGE_INPUT_TOKENS_CACHED_ATTRIBUTE, cachedInputTokens);
+
+        const cacheWriteInputTokens =
+          providerMetadataObject.anthropic.usage?.cache_creation_input_tokens ??
+          providerMetadataObject.anthropic.cacheCreationInputTokens;
+        setAttributeIfDefined(attributes, GEN_AI_USAGE_INPUT_TOKENS_CACHE_WRITE_ATTRIBUTE, cacheWriteInputTokens);
       }
 
       if (providerMetadataObject.bedrock?.usage) {
         setAttributeIfDefined(
           attributes,
-          'gen_ai.usage.input_tokens.cached',
+          GEN_AI_USAGE_INPUT_TOKENS_CACHED_ATTRIBUTE,
           providerMetadataObject.bedrock.usage.cacheReadInputTokens,
         );
         setAttributeIfDefined(
           attributes,
-          'gen_ai.usage.input_tokens.cache_write',
+          GEN_AI_USAGE_INPUT_TOKENS_CACHE_WRITE_ATTRIBUTE,
           providerMetadataObject.bedrock.usage.cacheWriteInputTokens,
         );
       }
 
       if (providerMetadataObject.deepseek) {
         setAttributeIfDefined(
           attributes,
-          'gen_ai.usage.input_tokens.cached',
+          GEN_AI_USAGE_INPUT_TOKENS_CACHED_ATTRIBUTE,
           providerMetadataObject.deepseek.promptCacheHitTokens,
         );
         setAttributeIfDefined(

packages/core/src/utils/vercel-ai/vercel-ai-attributes.ts

@@ -288,6 +288,14 @@ export const AI_RESPONSE_PROVIDER_METADATA_ATTRIBUTE = 'ai.response.providerMeta
  */
 export const AI_SETTINGS_MAX_RETRIES_ATTRIBUTE = 'ai.settings.maxRetries';
 
+/**
+ * Basic LLM span information
+ * Multiple spans
+ *
+ * The number of cached input tokens that were used
+ * @see https://ai-sdk.dev/docs/ai-sdk-core/telemetry#basic-llm-span-information
+ */
+export const AI_USAGE_CACHED_INPUT_TOKENS_ATTRIBUTE = 'ai.usage.cachedInputTokens';
 /**
  * Basic LLM span information
  * Multiple spans
@@ -863,6 +871,21 @@ interface AnthropicProviderMetadata {
    * @see https://ai-sdk.dev/providers/ai-sdk-providers/anthropic#cache-control
    */
   cacheReadInputTokens?: number;
+
+  /**
+   * Usage metrics for the Anthropic model.
+   */
+  usage?: {
+    input_tokens: number;
+    cache_creation_input_tokens?: number;
+    cache_read_input_tokens?: number;
+    cache_creation?: {
+      ephemeral_5m_input_tokens?: number;
+      ephemeral_1h_input_tokens?: number;
+    };
+    output_tokens?: number;
+    service_tier?: string;
+  };
 }
 
 /**