Hey, this is a known issue. When you set a custom OpenAI Base URL or API key, that config gets applied to all model providers, not just OpenAI.
The same thing was discussed here: After integrating the Open API key, the existing AI becomes unusable
Workaround: turn off the custom API settings (Settings > Models > OpenAI Base URL) before switching to Gemini or other native Cursor models.
The team is aware and the bug is logged. Your report will help with prioritization.