diff --git a/.changeset/stream-read-v3-reconnect.md b/.changeset/stream-read-v3-reconnect.md
new file mode 100644
index 0000000000..71fd0c843c
--- /dev/null
+++ b/.changeset/stream-read-v3-reconnect.md
@@ -0,0 +1,5 @@
+---
+'@workflow/world-vercel': patch
+---
+
+Use v3 endpoint for stream reads, which supports automatic transparent reconnects.
diff --git a/packages/world-vercel/src/streamer.test.ts b/packages/world-vercel/src/streamer.test.ts
index af44a6c755..515981b784 100644
--- a/packages/world-vercel/src/streamer.test.ts
+++ b/packages/world-vercel/src/streamer.test.ts
@@ -190,7 +190,7 @@ describe('streams.get', () => {
     vi.restoreAllMocks();
   });
 
-  it('includes runId in the fetch URL', async () => {
+  it('reads the live stream from the v3 endpoint (error-on-timeout)', async () => {
     const fetchSpy = vi
       .spyOn(globalThis, 'fetch')
       .mockImplementation(
@@ -202,10 +202,12 @@ describe('streams.get', () => {
 
     expect(fetchSpy).toHaveBeenCalledTimes(1);
     const url = new URL(fetchSpy.mock.calls[0][0] as string);
-    expect(url.pathname).toBe('/v2/runs/run-123/stream/my-stream');
+    // v3, not v2: the reconnecting reader relies on the server erroring the
+    // body on a max-duration timeout rather than closing it cleanly.
+    expect(url.pathname).toBe('/v3/runs/run-123/stream/my-stream');
   });
 
-  it('passes startIndex as a query parameter', async () => {
+  it('passes startIndex as a query parameter on the v3 read', async () => {
     const fetchSpy = vi
       .spyOn(globalThis, 'fetch')
       .mockImplementation(
@@ -216,7 +218,7 @@ describe('streams.get', () => {
     await streamer.streams.get('run-123', 'my-stream', 5);
 
     const url = new URL(fetchSpy.mock.calls[0][0] as string);
-    expect(url.pathname).toBe('/v2/runs/run-123/stream/my-stream');
+    expect(url.pathname).toBe('/v3/runs/run-123/stream/my-stream');
     expect(url.searchParams.get('startIndex')).toBe('5');
   });
 });
diff --git a/packages/world-vercel/src/streamer.ts b/packages/world-vercel/src/streamer.ts
index b74e2ae2e9..9be08bbe9a 100644
--- a/packages/world-vercel/src/streamer.ts
+++ b/packages/world-vercel/src/streamer.ts
@@ -23,12 +23,27 @@ export const MAX_CHUNKS_PER_REQUEST = 1000;
 // (partial writes, long-lived reads), and duplex streams are incompatible
 // with undici's experimental H2 support.
 
+// Writes (PUT) and stream completion use the v2 stream endpoint.
 function getStreamUrl(name: string, runId: string, httpConfig: HttpConfig) {
   return new URL(
     `${httpConfig.baseUrl}/v2/runs/${encodeURIComponent(runId)}/stream/${encodeURIComponent(name)}`
   );
 }
 
+// The live-read (GET) endpoint is versioned at v3: on a max-duration timeout
+// (or a mid-stream connection drop) the server errors the response body
+// instead of closing it cleanly, which is what lets the reconnecting reader
+// (`createReconnectingFramedStream`) resume from the next chunk rather than
+// treating the timeout as end-of-stream. Reading from v2 would silently
+// truncate long-lived streams at the server's 2-minute limit. Only the live
+// read is affected by the timeout — writes, completion, and snapshot reads
+// (chunks/info/list) stay on v2.
+function getStreamReadUrl(name: string, runId: string, httpConfig: HttpConfig) {
+  return new URL(
+    `${httpConfig.baseUrl}/v3/runs/${encodeURIComponent(runId)}/stream/${encodeURIComponent(name)}`
+  );
+}
+
 function createStreamRequestError(
   operation: 'write' | 'close',
   url: URL,
@@ -190,7 +205,7 @@ export function createStreamer(config?: APIConfig): Streamer {
 
       async get(runId: string, name: string, startIndex?: number) {
         const httpConfig = await getHttpConfig(config);
-        const url = getStreamUrl(name, runId, httpConfig);
+        const url = getStreamReadUrl(name, runId, httpConfig);
         if (typeof startIndex === 'number') {
           url.searchParams.set('startIndex', String(startIndex));
         }