Hey, thanks for the report. This is a known bug affecting GPT-5 High/Fast when using custom OpenAI API keys. Cursor is incorrectly sending the prompt_cache_retention parameter, which those models don’t support.
Temporary workaround: use GPT-5 Codex instead of GPT-5 High/Fast with your custom API key - it’s confirmed to work (see the related report: Custom API key breaks other models).
I’ll pass this to the team. Please let me know if the workaround resolves the issue.