diff --git a/packages/ai/src/route/transport/http.ts b/packages/ai/src/route/transport/http.ts
index c87f0d75c2d5..771df58b51a5 100644
--- a/packages/ai/src/route/transport/http.ts
+++ b/packages/ai/src/route/transport/http.ts
@@ -119,20 +119,27 @@ export const httpJson = <Body, Frame>(input: HttpJsonInput<Body, Frame>): HttpJs
       )
       const http = RequestExecutor.responseHttp(response)
       const remaining = Duration.subtract(total, Duration.millis((yield* Clock.currentTimeMillis) - started))
+      // SSE comment heartbeats (`: keepalive`) arrive as bytes but produce no frame, so for
+      // SSE the stall limit rides the framed stream — only real events re-arm it. Every other
+      // framing keeps byte-level progress.
+      const sse = prepared.framing.id === "sse"
+      const stall = Stream.timeoutOrElse({
+        duration: timeoutDuration(request.http?.chunkTimeout),
+        orElse: () => Stream.fail(timeout("read", "Timed out waiting for response data", http)),
+      })
       return {
-        frames: prepared.framing.frame(
-          RequestExecutor.responseStream(response).pipe(
-            Stream.timeoutOrElse({
-              duration: timeoutDuration(request.http?.chunkTimeout),
-              orElse: () => Stream.fail(timeout("read", "Timed out waiting for response data", http)),
-            }),
-            Stream.interruptWhen(
-              Effect.sleep(remaining).pipe(
-                Effect.andThen(Effect.fail(timeout("read", "Timed out waiting for the response to complete", http))),
+        frames: prepared.framing
+          .frame(
+            RequestExecutor.responseStream(response).pipe(
+              sse ? (stream) => stream : stall,
+              Stream.interruptWhen(
+                Effect.sleep(remaining).pipe(
+                  Effect.andThen(Effect.fail(timeout("read", "Timed out waiting for the response to complete", http))),
+                ),
               ),
             ),
-          ),
-        ),
+          )
+          .pipe(sse ? stall : (stream) => stream),
         http,
         body: prepared.framing.body,
       }
diff --git a/packages/ai/test/http-timeout.test.ts b/packages/ai/test/http-timeout.test.ts
index 6b45a8070079..03c70796d159 100644
--- a/packages/ai/test/http-timeout.test.ts
+++ b/packages/ai/test/http-timeout.test.ts
@@ -87,6 +87,38 @@ describe("HTTP transport timeouts", () => {
     }),
   )
 
+  // SSE comment heartbeats keep a stalled generation "warm" at the byte level without
+  // ever dispatching an event, so they must not extend the chunk timeout. Runs on the
+  // live clock so heartbeats arriving faster than the chunk timeout would keep a
+  // byte-level stall timer alive forever.
+  it.live("fails while a stalled stream only emits SSE comment heartbeats", () =>
+    Effect.gen(function* () {
+      const encoder = new TextEncoder()
+      const layer = dynamicResponse((input) =>
+        Effect.sync(() =>
+          input.respond(
+            new ReadableStream<Uint8Array>({
+              start(controller) {
+                controller.enqueue(encoder.encode(sseRaw(`data: ${JSON.stringify(deltaChunk({ content: "Hi" }))}`)))
+              },
+              async pull(controller) {
+                await new Promise((resolve) => setTimeout(resolve, 5))
+                controller.enqueue(encoder.encode(": keepalive\n\n"))
+              },
+            }),
+            SSE,
+          ),
+        ),
+      )
+      const error = yield* LLMClient.generate(
+        LLM.request({ model, prompt: "Hello", http: { chunkTimeout: 25 } }),
+      ).pipe(Effect.provide(layer), Effect.flip)
+
+      expect(error.reason).toMatchObject({ _tag: "Transport", transport: "http", operation: "read", code: "Timeout" })
+      expect(error.reason.http).toMatchObject({ status: 200 })
+    }),
+  )
+
   it.effect("applies a configured header timeout", () =>
     Effect.gen(function* () {
       const fiber = yield* LLMClient.generate(