Hey, thanks for the report. Good news: in the logs, the model finished all the work for this request successfully on the server. The client connection is what drops near the end of the turn. The Stream ended without turnEnded — connection likely dropped mid-stream error is transient and means the HTTP/2 stream closed before completion. On long, complex tasks with Fable 5 High, the stream stays open longer, so it’s more likely to get interrupted by a VPN, a corporate proxy, or an unstable network.
What to try:
If you’re on a VPN or behind a corporate proxy like Zscaler, open App Settings Cmd+,, find HTTP/2, and switch to HTTP/1.1. This is often the cause of drops on long streams.
Run Cursor Settings > Network > Run Diagnostics and check if it shows any clear connection issues.
Update Cursor. You’re on 3.10.20, and newer versions improved retry behavior when a stream drops.
Let me know if this helps, especially switching off HTTP/2. If the drops continue, please share a couple more Request IDs from fresh attempts.
Thanks for trying both steps and coming back with details. Since the disconnect is still happening on 3.11.19 with HTTP/1.1, the old Request ID won’t help anymore. It’s from a run before those changes.
Can you send a couple of fresh Request IDs from the attempts where it now “spins for a long time and then returns nothing” right top chat menu > Copy Request ID. If you have Privacy Mode enabled, the Request ID may be unavailable. In that case, please turn it off temporarily in Cursor Settings and try again so the ID gets saved.
A few more questions to narrow it down:
Are you still on a VPN or behind a corporate proxy like Zscaler? If yes, long streaming disconnects often come from that, even on HTTP/1.1.
What do you see in Cursor Settings > Network > Run Diagnostics?
Does the empty output happen on another model like Sonnet, or only on Fable 5 High?
With the new Request IDs, we can check the logs to see what’s happening on the server for those specific runs, whether the model finished or the stream is getting cut off early.
Thanks for the new Request ID and the answers, it really narrows things down.
The fact that it only drops on Fable 5 High, while GPT 5.6 sol and grok work fine, plus the corporate proxy and a 1 095 ms ping, is the key point. Fable 5 High produces longer thinking turns and keeps the connection open longer than the other models. With high latency behind a corporate proxy, that long stream is more likely to get cut off by the proxy idle timeout, even on HTTP/1.1. The other models respond faster, so they don’t hit the timeout.
What to try:
Test Fable 5 High on a network without the corporate proxy, for example a personal hotspot. If it doesn’t drop there, it’s almost certainly the proxy idle timeout. Then you can ask IT to increase the timeout for long streaming connections to the Cursor domain.
A 1 095 ms ping is pretty high. If you can switch to a different network path or location, that can also help long streams.
As a temporary workaround for complex tasks behind the proxy, use Fable 5 without High or the models that already work reliably for you.
If you test on a network without the proxy, let me know whether it still drops or not. That will confirm the cause right away.