I’m giving Cursor a new feature to whack out and towards the end I tell it to fix some simple unit tests - it keeps going in circles, creating more and more elaborate “solutions” to try and fix the tests until I realise Composer-1 is selected.
I change to GPT-5.2 and ask it wtf it’s doing and question it’s last decision - it immediately undid all of Composer-1’s work and fixed the problem. I’m guessing Opus would have given the same result.
Meanwhile, I’m burning through usage with Composer-1 - is this a convenient business model for Cursor or something?
The team is aware of an issue where Composer-1 can get stuck in a loop in certain scenarios, and we’re actively working on a fix.
You can change the default model in the agent settings. Just pick the model you want (GPT-5.2, Sonnet 4.5, or another one) from the dropdown before you start. Your choice will be saved for new chats.
About usage cost: Composer-1 really can be more expensive when it loops. I’d recommend using other models for now until the issue is fixed.
If you want, you can share more details about the task where the looping happened. That would help the team debug it.