So I was using Auto mode for research tasks and at some point Cursor decided that it’s going to use claude-4.6-opus-high-thinking for all subagents / skills, etc.
As a result a single prompt used about 300M of API tokens (essentially 70% of my monthly Opus quota).
I thought by selecting Auto I was using Auto + composer (which seemed to be the case until yesterday), yet turns out if you pick Auto you can be running claude-4.6-opus-high-thinking.
This is what I got from Cursor’s support bot:
How to prevent this in the future:
To maintain control over which model is used and avoid unexpected costs:
- Turn off “Auto-select” in the model menu
- Manually select your preferred model (e.g., Claude 4.5 Sonnet, Gemini, GPT-5)
- Avoid using Auto mode for long-running or complex tasks unless you’re comfortable with it potentially routing to expensive models like Opus
Did anyone get confused like me about this? Did you run into the issue of Auto choosing expensive models (and eating up 70% of your Ultra subscription with a single prompt)? Was this always the case or it’s something newly introduced? What do you think about the recent division by API vs Auto + Composer in the recent billing changes?
The key takeaway for me now is that ‘Auto’ mode does not mean ‘low-cost’ or ‘Composer’ mode, it means most powerful available for the task, which in my case was the most expensive.
Staying away from the Auto mode for now until I figure out how the tokens get used. Mapping the token spending to my cursor subscription has become a real pain point in the last couple of months.
