diff --git a/dev-packages/node-integration-tests/suites/tracing/vercelai/v6_v7/scenario-cache-tokens.mjs b/dev-packages/node-integration-tests/suites/tracing/vercelai/v6_v7/scenario-cache-tokens.mjs
new file mode 100644
index 000000000000..6e8d12a81fac
--- /dev/null
+++ b/dev-packages/node-integration-tests/suites/tracing/vercelai/v6_v7/scenario-cache-tokens.mjs
@@ -0,0 +1,31 @@
+import * as Sentry from '@sentry/node';
+import { generateText } from 'ai';
+import { MockLanguageModelV3 } from 'ai/test';
+
+// Calls routed through the Vercel AI Gateway carry a `gateway` key in `providerMetadata` instead of a
+// provider name, so cache counts can only be derived from the SDK's normalized usage object.
+async function run() {
+  await Sentry.startSpan({ op: 'function', name: 'main' }, async () => {
+    await generateText({
+      experimental_telemetry: { isEnabled: true, recordInputs: true, recordOutputs: true },
+      model: new MockLanguageModelV3({
+        doGenerate: async () => ({
+          finishReason: { unified: 'stop', raw: 'stop' },
+          usage: {
+            inputTokens: { total: 120, noCache: 20, cacheRead: 80, cacheWrite: 20 },
+            outputTokens: { total: 10, noCache: 10, cached: 0 },
+            totalTokens: { total: 130, noCache: 30, cached: 100 },
+          },
+          content: [{ type: 'text', text: 'Cache token span!' }],
+          warnings: [],
+          providerMetadata: {
+            gateway: { routing: {} },
+          },
+        }),
+      }),
+      messages: [{ role: 'user', content: 'Tell me something' }],
+    });
+  });
+}
+
+run();
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 5c538267e935..d014e9308862 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
@@ -15,6 +15,7 @@ import {
   GEN_AI_TOOL_DEFINITIONS,
   GEN_AI_TOOL_DESCRIPTION,
   GEN_AI_TOOL_NAME,
+  GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS,
   GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS,
   GEN_AI_USAGE_INPUT_TOKENS,
   GEN_AI_USAGE_OUTPUT_TOKENS,
@@ -811,6 +812,43 @@ describe.each(matrix)('Vercel AI integration (version %s)', (version, vercelAiVe
     },
   );
 
+  createEsmTests(
+    __dirname,
+    'scenario-cache-tokens.mjs',
+    'instrument.mjs',
+    (createRunner, test) => {
+      test('reads cache token counts from the SDK usage object', async () => {
+        await createRunner()
+          .expect({ transaction: { transaction: 'main' } })
+          .expect({
+            span: container => {
+              const generateContent = container.items.find(
+                span => span.attributes['sentry.op']?.value === 'gen_ai.generate_content',
+              )!;
+              expect(generateContent).toBeDefined();
+              expect(generateContent.attributes[GEN_AI_USAGE_INPUT_TOKENS]?.value).toBe(120);
+              expect(generateContent.attributes[GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS]?.value).toBe(80);
+              expect(generateContent.attributes[GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS]?.value).toBe(20);
+
+              const invokeAgent = container.items.find(
+                span => span.attributes['sentry.op']?.value === 'gen_ai.invoke_agent',
+              )!;
+              expect(invokeAgent).toBeDefined();
+              expect(invokeAgent.attributes[GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS]?.value).toBe(80);
+              expect(invokeAgent.attributes[GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS]?.value).toBe(20);
+            },
+          })
+          .start()
+          .completed();
+      });
+    },
+    {
+      additionalDependencies: {
+        ai: vercelAiVersion,
+      },
+    },
+  );
+
   createEsmTests(
     __dirname,
     'scenario-embeddings.mjs',