Desktop model catalog filtered since June 29, but web works (same account)

@soyking_k, yeah, you’re right. disableHttp2 by itself doesn’t fix this. The issue is that the local process that fetches the model catalog uses its own HTTP/2 network stack, and it doesn’t pick up http.proxy, the system proxy, env vars, or disableHttp2. So the request goes out directly from a CN IP, and the catalog gets filtered by region. On 3.7.42 this request went over HTTP/1.1 through the system proxy, which is why the behavior differs between versions.

This behavior is known, and honestly I don’t have a clean application-layer workaround that works reliably on the current version right now. A TUN setup intercepts traffic at the network level, but it conflicts with Tailscale and corporate VPN, as mentioned above.

Thanks everyone for the detailed digging in the thread, especially for lsof. It really helped pinpoint what’s happening.

This bug affects BYOK usage as well.
Sometimes I use my own OpenAI API key, yet I am unable to add or use the GPT-5.5 model from the model list. If you don’t fix this soon, I think it might make me give up on using Cursor.

Does Cursor not plan to fix this issue?

Hey @Suilta_Pico, thanks for calling out BYOK. That’s an important detail.

It looks like the same thing we discussed earlier in the thread is happening here: the model catalog is filtered on the server side based on the request’s IP region, and the local process that fetches the model list currently isn’t going through the proxy. It uses its own HTTP/2 stack, which doesn’t pick up http.proxy, the system proxy, env vars, or disableHttp2. So if GPT-5.5 is filtered out of the catalog for your region, you can’t add it even if you’re using your own OpenAI key, because the model just never shows up in the picker.

I agree this feels weird for BYOK. If you’re using your own key, you’d expect the catalog not to be region-limited. I want to verify this separately so we don’t mix it with the main case. Can you share:

  • your Cursor version and OS
  • the Request ID (top right of the chat > Copy Request ID), if you see one when you try to use the model
  • a screenshot of the step where you try to add or select GPT-5.5 with BYOK enabled

If we confirm this is separate from the regional filtering behavior, I’ll suggest opening a separate thread for it so it doesn’t get lost in this one.

On the proxy behavior itself, the local process not respecting the proxy on 3.8+ is a known limitation and I’ve passed it to the team. I can’t share an exact date yet, but once I have something concrete, I’ll update this thread.

the same issue !!!
latest version, it’s not working. the 3.7x version , it can see all models. i don’t know why, but cursor should fix the issue.