diff --git a/packages/server-utils/src/integrations/vercel-ai/vercel-ai-dc-subscriber.ts b/packages/server-utils/src/integrations/vercel-ai/vercel-ai-dc-subscriber.ts
index c1daaf73fc84..2e1c173ee08e 100644
--- a/packages/server-utils/src/integrations/vercel-ai/vercel-ai-dc-subscriber.ts
+++ b/packages/server-utils/src/integrations/vercel-ai/vercel-ai-dc-subscriber.ts
@@ -951,9 +951,9 @@ function buildInputMessageAttributes(event: Record<string, unknown>): Record<str
   // `ai` >= 7 forbids system messages in `messages`/`prompt` and exposes the system prompt as a
   // separate `instructions` field. The OTel path lifts the system message out of the prompt into
   // `gen_ai.system_instructions` as `[{ type: 'text', content }]`; mirror that shape here.
-  const instructions = asString(event.instructions);
+  const instructions = systemInstructionEntries(event.instructions);
   if (instructions) {
-    attributes[GEN_AI_SYSTEM_INSTRUCTIONS] = stringify([{ type: 'text', content: instructions }]);
+    attributes[GEN_AI_SYSTEM_INSTRUCTIONS] = stringify(instructions);
   }
 
   // The AI SDK start events extend `StandardizedPrompt`; messages live on `messages`, otherwise the
@@ -965,3 +965,22 @@ function buildInputMessageAttributes(event: Record<string, unknown>): Record<str
 
   return attributes;
 }
+
+/**
+ * Normalize the `system`/`instructions` option into `gen_ai.system_instructions` entries. The AI SDK
+ * accepts `string | SystemModelMessage | SystemModelMessage[]` there — the message form is how a
+ * caller attaches `providerOptions` to the system prompt (e.g. Anthropic prompt caching) — so a bare
+ * string and single message both become one entry, and an array contributes one entry per message.
+ */
+function systemInstructionEntries(value: unknown): Array<{ type: 'text'; content: string }> | undefined {
+  if (typeof value === 'string') {
+    return [{ type: 'text', content: value }];
+  }
+  const messages = Array.isArray(value) ? value : isObjectLike(value) ? [value] : [];
+  const entries = messages.filter(isObjectLike).flatMap(message => {
+    const content = asString(message.content);
+    return content ? [{ type: 'text' as const, content }] : [];
+  });
+
+  return entries.length ? entries : undefined;
+}
diff --git a/packages/server-utils/src/integrations/vercel-ai/vercel-ai-orchestrion-subscriber.ts b/packages/server-utils/src/integrations/vercel-ai/vercel-ai-orchestrion-subscriber.ts
index 2a59de79b4f8..a8b8c770feb1 100644
--- a/packages/server-utils/src/integrations/vercel-ai/vercel-ai-orchestrion-subscriber.ts
+++ b/packages/server-utils/src/integrations/vercel-ai/vercel-ai-orchestrion-subscriber.ts
@@ -663,10 +663,12 @@ function buildTextMessage(type: 'generateText' | 'streamText' | 'generateObject'
       functionId: asString(telemetry.functionId),
       ...modelFields(options.model),
       maxRetries: options.maxRetries,
-      // The `ai` SDK takes the system prompt as a top-level `system` option (all of v4/v5/v6); the
-      // shared core lifts `event.instructions` into the system-instructions attribute, matching v7's
-      // native channel (which carries it as a distinct field rather than inside the messages array).
-      instructions: asString(options.system),
+      // The `ai` SDK takes the system prompt as a top-level `system` option (all of v4/v5/v6) that can
+      // be a string, a `SystemModelMessage`, or an array of them; the shared core lifts
+      // `event.instructions` into the system-instructions attribute, matching v7's native channel
+      // (which carries it as a distinct field rather than inside the messages array), so pass it
+      // through unchanged rather than narrowing it to a string here.
+      instructions: options.system,
       // Normalize to the message-array shape the shared core (and v7's channel) expects: a bare string
       // `prompt` becomes a single user message, matching the SDK's own normalization.
       messages: normalizePromptMessages(options),
diff --git a/packages/server-utils/test/integrations/vercel-ai/system-instructions.test.ts b/packages/server-utils/test/integrations/vercel-ai/system-instructions.test.ts
new file mode 100644
index 000000000000..cee02db1261a
--- /dev/null
+++ b/packages/server-utils/test/integrations/vercel-ai/system-instructions.test.ts
@@ -0,0 +1,70 @@
+import { afterEach, beforeEach, describe, expect, it } from 'vitest';
+import { GEN_AI_SYSTEM_INSTRUCTIONS } from '@sentry/conventions/attributes';
+import { getMainCarrier, setCurrentClient, spanToStaticSpanJSON } from '@sentry/core';
+import type { Span } from '@sentry/core';
+import { createSpanFromMessage } from '../../../src/integrations/vercel-ai/vercel-ai-dc-subscriber';
+import { getDefaultTestClientOptions, TestClient } from '../../mocks/client';
+
+describe('Vercel AI system instructions', () => {
+  beforeEach(() => {
+    getMainCarrier().__SENTRY__ = undefined;
+  });
+
+  afterEach(() => {
+    getMainCarrier().__SENTRY__ = undefined;
+  });
+
+  function setupClient(): Span[] {
+    const client = new TestClient(
+      getDefaultTestClientOptions({
+        dsn: 'https://public@dsn.ingest.sentry.io/1337',
+        tracesSampleRate: 1,
+      }),
+    );
+    setCurrentClient(client);
+    client.init();
+
+    const endedSpans: Span[] = [];
+    client.on('spanEnd', span => endedSpans.push(span));
+    return endedSpans;
+  }
+
+  it('records `gen_ai.system_instructions` when `instructions` is a string', () => {
+    const endedSpans = setupClient();
+    const span = createSpanFromMessage(
+      {
+        type: 'generateText',
+        event: { instructions: 'Be brief.', recordInputs: true },
+      } as Parameters<typeof createSpanFromMessage>[0],
+      {} as Parameters<typeof createSpanFromMessage>[1],
+    );
+    span?.end();
+
+    const data = spanToStaticSpanJSON(endedSpans[0]!).data;
+    expect(data?.[GEN_AI_SYSTEM_INSTRUCTIONS]).toBe('[{"type":"text","content":"Be brief."}]');
+  });
+
+  // AI SDK v7's `instructions` option (and v4-v6's `system`) accepts a `SystemModelMessage` object, not
+  // just a plain string -- for example to attach `providerOptions` like Anthropic prompt caching.
+  it('records `gen_ai.system_instructions` when `instructions` is a message object', () => {
+    const endedSpans = setupClient();
+    const span = createSpanFromMessage(
+      {
+        type: 'generateText',
+        event: {
+          instructions: {
+            role: 'system',
+            content: 'Be brief.',
+            providerOptions: { anthropic: { cacheControl: { type: 'ephemeral' } } },
+          },
+          recordInputs: true,
+        },
+      } as Parameters<typeof createSpanFromMessage>[0],
+      {} as Parameters<typeof createSpanFromMessage>[1],
+    );
+    span?.end();
+
+    const data = spanToStaticSpanJSON(endedSpans[0]!).data;
+    expect(data?.[GEN_AI_SYSTEM_INSTRUCTIONS]).toBe('[{"type":"text","content":"Be brief."}]');
+  });
+});