I’m on the Pro plan (account: Oleksii Patrykieiev). Almost all models fail with a connection error. This has been happening from the beginning, not just recently. Usage limits are not exhausted.
What works
Auto — works
Composer — works
What fails
Cursor Grok (any variant)
Claude (including Claude Sonnet 4.6 / 4.5, also 1M)
GPT and other third-party models
Exact error
In the chat I get:
Network Error
We’re having trouble connecting to the model provider. This might be temporary — please try again in a moment.
When I dig deeper, the response is not a normal model reply — it’s a Cloudflare block page (HTML with “Why have I been blocked?”, Cloudflare Ray ID, etc.). So the request seems blocked by Cloudflare on the way to the model provider.
Usage (Billing / Included Usage) — cycle Sep 9 – Oct 9, 2026
Cursor Models: ~49.6% used (mostly auto ~46.9%)
Other Models: ~15.3% used (claude-3.5-sonnet earlier)
On-demand: $0.00 (I also tried raising the fixed limit to $10 — did not help)
So this is not a quota or billing issue.
What I already tried
Switching models (Claude without 1M, GPT, Grok, etc.)
New chats / retry
Raising on-demand monthly limit
Checking that included usage is still available
Nothing fixed Claude / Grok / GPT. Only Auto and Composer keep working.
Ask
Please check why my requests are getting a Cloudflare block for non-Composer models, and restore access to Grok / Claude / GPT on my Pro account.
I’m too busy to write a detailed bug report, but I’m having similar issues. Agents are just dropping and I can’t reconnect to them even over web. It’s totally random but seems to be the Grok agents mostly.
First, the good news. Your account and usage look fine, and Cursor isn’t being blocked by Cloudflare. The issue is a setting on your side.
It looks like you have a personal API key enabled (BYOK), most likely Use Azure OpenAI with an empty Base URL and deployment. When any personal key toggle is enabled, all models except Cursor’s own models go through that key config instead of Cursor. That’s why Claude, GPT, and Grok fail. The Cloudflare page is what that endpoint returns when the request is made with a non-Cursor key. Auto and Composer keep working because they always run on Cursor infra.
How to fix it:
Open Cursor Settings (gear icon in the top right) > Models.
Scroll to the API Keys section.
Turn off Use Azure OpenAI.
Also make sure Use OpenAI API Key, Use Anthropic API Key, Use Google API Key, and Use AWS Bedrock are off, and clear Override OpenAI Base URL if anything is set there.
Open a new chat and try Claude Sonnet 4.6, then Grok.
Models should start working right away and usage should count against your Pro included usage as usual.
One more thing. Showing a generic Network Error instead of a clear config message is not great. I’ve passed this to the team so validation is clearer. If anything still fails after step 5, send the Request ID next to the error in the chat, and I’ll take another look.
Turning off the personal API key toggles (using the steps above) is the fix for your case, so Claude, GPT, and Grok should’ve started working right after that. Based on our data, that’s exactly what happened.
Also, we fixed the issue where Grok showed a raw Cloudflare page when the personal API key was enabled. The change is rolling out now. So even if you accidentally leave the toggle on, you’ll see clearer behavior instead of a confusing block.
If anything still isn’t working after changing the settings, send the Request ID shown next to the error in the chat, and I’ll take another look.
Yes, I corrected the settings and I also see that you fixed things on your side. Everything works great. Thanks for the help. I am testing it, but I think everything will be fine, and if anything comes up, I’ll let you know. Giving a thumbs up to your support, thank you.