Hey, thanks for the report.
This is a known issue: when using BYOK (Override OpenAI Base URL) with GPT models, Cursor sends requests in the Responses API format (with the input field) instead of the Chat Completions API format (with the messages field). OpenAI-compatible providers expect the standard format and return an error.
The team is working on a fix, but there’s no ETA yet.
Workarounds:
- Turn off BYOK for GPT models and use Cursor’s built-in API (Settings > Models, remove the OpenAI API Key and Override Base URL)
- Use a proxy that can convert the Responses API payload into the Chat Completions format (for example, LiteLLM proxy)
Related thread with the same issue: Requests are sent to incorrect endpoint when using base URL override