Hey, thanks for the detailed report, and yep, you’re right. A generic “by design” answer doesn’t fit here, because you’re talking about routing logic, not whether Composer 2 can run on the OpenAI endpoint.
Confirmed on the substance:
- The current behavior is a routing choice, not a technical limit. The backend rejects requests to Composer models if any BYOK key is present in settings. In Auto mode, the key is automatically removed on the backend and the request goes through Cursor. For an explicitly selected Composer model, that doesn’t happen yet, which is why you see the error.
- The per-model routing you described (OpenAI key only for OpenAI models, Anthropic key only for Anthropic, proprietary models always via Cursor) is the right direction. This is tracked as a feature request: Auto-route between API key models and Cursor plan models without toggling settings. I can’t share an ETA.
- We’ve logged the issue on our side. I’ll pass your report and your point 2 reasoning along as extra context for the team.
Workaround for now (I know this is the problem, not a real fix): temporarily disable the OpenAI key in Settings when you need Composer 2. I get that this is exactly the UX you’re calling out, I just want you to have something that unblocks you today.
If there’s an update on the routing logic, I’ll reply in the thread.