diff --git a/MIGRATION.md b/MIGRATION.md
index 9582bf374888..da52af69448c 100644
--- a/MIGRATION.md
+++ b/MIGRATION.md
@@ -256,6 +256,7 @@ Affected SDKs: All SDKs.
 - The `gen_ai.request.available_tools` span attribute was renamed to `gen_ai.tool.definitions` across all AI integrations.
 - The `gen_ai.tool.input` span attribute was renamed to `gen_ai.tool.call.arguments` across all AI integrations.
 - The `gen_ai.tool.output` span attribute was renamed to `gen_ai.tool.call.result` across all AI integrations.
+- The Vercel AI token attributes `gen_ai.usage.input_tokens.cached`, `gen_ai.usage.input_tokens.cache_write`, and `gen_ai.usage.output_tokens.reasoning` were renamed to `gen_ai.usage.cache_read.input_tokens`, `gen_ai.usage.cache_creation.input_tokens`, and `gen_ai.usage.reasoning.output_tokens`.
 - Span attributes now use the shared `@sentry/conventions` package under the hood.
 
 If you reference these attributes in custom instrumentation, `beforeSendSpan`, dashboards, or alerts, update them to the new names.
diff --git a/dev-packages/node-integration-tests/suites/tracing/vercelai/v6_v7/test.ts b/dev-packages/node-integration-tests/suites/tracing/vercelai/v6_v7/test.ts
index 17ca077a0422..6aa2280e57a3 100644
--- a/dev-packages/node-integration-tests/suites/tracing/vercelai/v6_v7/test.ts
+++ b/dev-packages/node-integration-tests/suites/tracing/vercelai/v6_v7/test.ts
@@ -16,9 +16,10 @@ import {
   GEN_AI_TOOL_DESCRIPTION,
   GEN_AI_TOOL_NAME,
   GEN_AI_TOOL_TYPE,
+  GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS,
   GEN_AI_USAGE_INPUT_TOKENS,
-  GEN_AI_USAGE_INPUT_TOKENS_CACHED,
   GEN_AI_USAGE_OUTPUT_TOKENS,
+  GEN_AI_USAGE_REASONING_OUTPUT_TOKENS,
   GEN_AI_USAGE_TOTAL_TOKENS,
 } from '@sentry/conventions/attributes';
 import { GEN_AI_TOOL_CALL_ID_ATTRIBUTE } from '../../../../../../packages/core/src/tracing/ai/gen-ai-attributes';
@@ -793,8 +794,8 @@ describe.each(matrix)('Vercel AI integration (version %s)', (version, vercelAiVe
               // Cache/reasoning token breakdown and conversation id are derived from the model's
               // `providerMetadata` — by the OTel processor on v6 and by the channel subscriber on v7,
               // both via the shared `getProviderMetadataAttributes` helper, so the shape is identical.
-              expect(generateContent.attributes[GEN_AI_USAGE_INPUT_TOKENS_CACHED]?.value).toBe(5);
-              expect(generateContent.attributes['gen_ai.usage.output_tokens.reasoning']?.value).toBe(7);
+              expect(generateContent.attributes[GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS]?.value).toBe(5);
+              expect(generateContent.attributes[GEN_AI_USAGE_REASONING_OUTPUT_TOKENS]?.value).toBe(7);
               expect(generateContent.attributes[GEN_AI_CONVERSATION_ID]?.value).toBe('resp_abc123');
 
               const invokeAgent = container.items.find(
diff --git a/packages/core/src/tracing/vercel-ai/index.ts b/packages/core/src/tracing/vercel-ai/index.ts
index 12a02030db3e..0ac89af81399 100644
--- a/packages/core/src/tracing/vercel-ai/index.ts
+++ b/packages/core/src/tracing/vercel-ai/index.ts
@@ -24,10 +24,11 @@ import {
   GEN_AI_TOOL_DESCRIPTION,
   GEN_AI_TOOL_NAME,
   GEN_AI_TOOL_TYPE,
+  GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS,
+  GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS,
   GEN_AI_USAGE_INPUT_TOKENS,
-  GEN_AI_USAGE_INPUT_TOKENS_CACHED,
-  GEN_AI_USAGE_INPUT_TOKENS_CACHE_WRITE,
   GEN_AI_USAGE_OUTPUT_TOKENS,
+  GEN_AI_USAGE_REASONING_OUTPUT_TOKENS,
   GEN_AI_USAGE_TOTAL_TOKENS,
 } from '@sentry/conventions/attributes';
 import { GEN_AI_TOOL_CALL_ID_ATTRIBUTE } from '../ai/gen-ai-attributes';
@@ -256,7 +257,7 @@ function buildOutputMessages(attributes: Record<string, unknown>): void {
 export function processVercelAiSpanAttributes(attributes: Record<string, unknown>): void {
   renameAttributeKey(attributes, AI_USAGE_COMPLETION_TOKENS_ATTRIBUTE, GEN_AI_USAGE_OUTPUT_TOKENS);
   renameAttributeKey(attributes, AI_USAGE_PROMPT_TOKENS_ATTRIBUTE, GEN_AI_USAGE_INPUT_TOKENS);
-  renameAttributeKey(attributes, AI_USAGE_CACHED_INPUT_TOKENS_ATTRIBUTE, GEN_AI_USAGE_INPUT_TOKENS_CACHED);
+  renameAttributeKey(attributes, AI_USAGE_CACHED_INPUT_TOKENS_ATTRIBUTE, GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS);
 
   // Parent spans (ai.streamText, ai.streamObject, etc.) use inputTokens/outputTokens instead of promptTokens/completionTokens
   renameAttributeKey(attributes, 'ai.usage.inputTokens', GEN_AI_USAGE_INPUT_TOKENS);
@@ -276,10 +277,10 @@ export function processVercelAiSpanAttributes(attributes: Record<string, unknown
   if (
     !inputTokensAreCacheInclusive &&
     typeof attributes[GEN_AI_USAGE_INPUT_TOKENS] === 'number' &&
-    typeof attributes[GEN_AI_USAGE_INPUT_TOKENS_CACHED] === 'number'
+    typeof attributes[GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS] === 'number'
   ) {
     attributes[GEN_AI_USAGE_INPUT_TOKENS] =
-      attributes[GEN_AI_USAGE_INPUT_TOKENS] + attributes[GEN_AI_USAGE_INPUT_TOKENS_CACHED];
+      attributes[GEN_AI_USAGE_INPUT_TOKENS] + attributes[GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS];
   }
 
   // Compute total tokens from input + output (embeddings may only have input tokens)
@@ -546,8 +547,8 @@ export function getProviderMetadataAttributes(providerMetadata: unknown): Record
   // OpenAI (v5 uses 'openai', v6 Azure Responses API uses 'azure')
   const openaiMetadata: OpenAiProviderMetadata | undefined = metadata.openai ?? metadata.azure;
   if (openaiMetadata) {
-    setAttributeIfDefined(attributes, GEN_AI_USAGE_INPUT_TOKENS_CACHED, openaiMetadata.cachedPromptTokens);
-    setAttributeIfDefined(attributes, 'gen_ai.usage.output_tokens.reasoning', openaiMetadata.reasoningTokens);
+    setAttributeIfDefined(attributes, GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, openaiMetadata.cachedPromptTokens);
+    setAttributeIfDefined(attributes, GEN_AI_USAGE_REASONING_OUTPUT_TOKENS, openaiMetadata.reasoningTokens);
     setAttributeIfDefined(
       attributes,
       'gen_ai.usage.output_tokens.prediction_accepted',
@@ -564,24 +565,28 @@ export function getProviderMetadataAttributes(providerMetadata: unknown): Record
   if (metadata.anthropic) {
     const cachedInputTokens =
       metadata.anthropic.usage?.cache_read_input_tokens ?? metadata.anthropic.cacheReadInputTokens;
-    setAttributeIfDefined(attributes, GEN_AI_USAGE_INPUT_TOKENS_CACHED, cachedInputTokens);
+    setAttributeIfDefined(attributes, GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, cachedInputTokens);
 
     const cacheWriteInputTokens =
       metadata.anthropic.usage?.cache_creation_input_tokens ?? metadata.anthropic.cacheCreationInputTokens;
-    setAttributeIfDefined(attributes, GEN_AI_USAGE_INPUT_TOKENS_CACHE_WRITE, cacheWriteInputTokens);
+    setAttributeIfDefined(attributes, GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS, cacheWriteInputTokens);
   }
 
   if (metadata.bedrock?.usage) {
-    setAttributeIfDefined(attributes, GEN_AI_USAGE_INPUT_TOKENS_CACHED, metadata.bedrock.usage.cacheReadInputTokens);
     setAttributeIfDefined(
       attributes,
-      GEN_AI_USAGE_INPUT_TOKENS_CACHE_WRITE,
+      GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS,
+      metadata.bedrock.usage.cacheReadInputTokens,
+    );
+    setAttributeIfDefined(
+      attributes,
+      GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS,
       metadata.bedrock.usage.cacheWriteInputTokens,
     );
   }
 
   if (metadata.deepseek) {
-    setAttributeIfDefined(attributes, GEN_AI_USAGE_INPUT_TOKENS_CACHED, metadata.deepseek.promptCacheHitTokens);
+    setAttributeIfDefined(attributes, GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, metadata.deepseek.promptCacheHitTokens);
     setAttributeIfDefined(attributes, 'gen_ai.usage.input_tokens.cache_miss', metadata.deepseek.promptCacheMissTokens);
   }
 
diff --git a/packages/core/test/lib/tracing/vercel-ai-cached-tokens.test.ts b/packages/core/test/lib/tracing/vercel-ai-cached-tokens.test.ts
index 4ab68c6f8d1b..59f8884fff66 100644
--- a/packages/core/test/lib/tracing/vercel-ai-cached-tokens.test.ts
+++ b/packages/core/test/lib/tracing/vercel-ai-cached-tokens.test.ts
@@ -38,7 +38,7 @@ describe('vercel-ai cached tokens', () => {
     });
 
     expect(span.data?.['gen_ai.usage.input_tokens']).toBe(150);
-    expect(span.data?.['gen_ai.usage.input_tokens.cached']).toBe(50);
+    expect(span.data?.['gen_ai.usage.cache_read.input_tokens']).toBe(50);
     expect(span.data?.['gen_ai.usage.total_tokens']).toBe(150);
   });
 
@@ -57,7 +57,7 @@ describe('vercel-ai cached tokens', () => {
     });
 
     expect(span.data?.['gen_ai.usage.input_tokens']).toBe(9500);
-    expect(span.data?.['gen_ai.usage.input_tokens.cached']).toBe(8000);
+    expect(span.data?.['gen_ai.usage.cache_read.input_tokens']).toBe(8000);
     expect(span.data?.['gen_ai.usage.output_tokens']).toBe(300);
     expect(span.data?.['gen_ai.usage.total_tokens']).toBe(9800);
   });