I want to say something because it concerns an executive decision Cursor made recently with its new model.
Cursor 1.5 was an epiphany! It showed perfect understanding, it worked thoroughly through any task at hand, and delivered what it was asked to deliver. It examined extensively codebase and showed concern and will to provide grounded answers and solutions.
Cursor 2 is just an average approximation. I understand that it is cheaper, and that strategically an expensive model is not a good choice for a vendor model, but… It is not like 1.5, and definitely not better.
Cursor 2 is spending a ridiculously large amount of tokens in thinking garbage (thankfully we can only see a few things from this). It also has a lot of attitude - yes in some cases it is for safety, which is not always bad, but for matters of architecture it gets on the way and I have to change model just to proceed with implementation of my directions.
I do a ton of iteration rather then really big plans and tasks. When i first started using composer 2 it felt like it had not been as good. I’ve started using it for more and more and even planning. Its really hard to say its much better then composer 1.5 but its still quite good.
I’ve had claude models give broken changes more lately.
I use convex + nextjs a lot so i think this may be highly dependent on how good your existing DX is.
The team is actively collecting feedback on Composer 2 in the official thread: Share your experience with Composer 2!. If you have specific examples where Composer 2 performs worse than 1.5 like prompt types or the kind of pushback it gives, it’d be really helpful to share them there. The more specific the examples, the easier it is for the team to prioritize improvements.
About the “attitude” on architecture decisions, for now the workaround is what you’re already doing, switch to another model for those tasks. In agent mode you can choose Claude or GPT for requests that need strict instruction following.
Composer 1.5 was replaced by Composer 2 Fast, so going back isn’t possible. But your feedback is being tracked, and the team knows some users felt 1.5 worked better for them.