then keep it free till you fix it
Hi team — I’ve recently started using Composer-2 and have been really impressed with the quality of the output so far 
I had a small suggestion that might help improve results in certain cases. It seems like Composer may not always be aware of the current year, which can impact web/RAG queries. For example, I’ve noticed it sometimes searches for things like “API endpoints for 2025,” which can lead to slightly outdated results.
It might be helpful to anchor queries to the current year (or dynamically adjust them) to keep results as fresh and relevant as possible.
Just a thought — overall really enjoying the experience 
I’ve just created an account for this. Can’t we use the Composer 1.5 version instead of Composer 2? It keeps upgrading me to Composer 2, but it is unusable; it generates lots of code I don’t want. I tried using an external model through the API option, but Cursor keeps overwhelming the context with too many tokens and then times out.
Composer 2 over-engineers everything; it doesn’t take into account existing coding patterns on the project and keeps duplicating functions everywhere for things that were already solved elsewhere in the code. I had a better experience using Composer 1.5 than 2. I’m spending money on Anthropic and hate it! I am trying to support Cursor, but this Composer 2 model has been a downgrade for me.
1 Like
I’ve been having massive issues with composure 2 destroying my code base. I have spent the past few days building a highly tuned python vision library with high speed cameras. And every time I ask for help with something it makes the worst mistakes.
Issue one - it’s ignoring my utility classes and building garbage versions in ui classes, or unrelated classes instead of using the high speed classes I’ve optimized.
Issue two - I ask for a simple change, and it makes 50+ edits then pushing to git? It pushed a mountain of garbage non compilable code to git.
Issue three - I was on a tight deadline for a contract so I asked it to fix a broken git head while I was working on the project. It deleted the local and deleted the remote branch, and then did a disk whip and entirely delete the files? Yeah so that lost me the client and the contract which was about two months of work and two months of pay. That left me screwed.
I’m on the 200 a month plan, I work in research and development and every single month I run out of usage credits with working back and forth with the AI. Claude sonnet 4.6 is the best one but I’m our of credits so easy then it switched to composure and I’m left with an AI that’s a ticking time bomb.
Don’t use composer 2, this so was not ready to be put out, they need to bring back 1.5 or give out more credits and take this AI down. This AI is doing real world damage to people’s projects, and lively hoods. If I had to choose between it or copy pasting code back and forth with a chat bot, I would do the copy pasting or just writing all the code myself. But then what’s the point of using cursor. Take my money. Destroy my work, and lose me clients, y’all need to do better man.
4 Likes
For me, every time Opus 4.5 craps itself (happening more frequently), Composer 2 seems to fix the problem. I agree with others that instructions need to be very clear. I have an ongoing side chat with GPT, which takes my prompts and makes them composer 2 foolproof. Audit and security output from Composer 2 shows results similar to those of Opus 4.5, with Opus being more comprehensive. Cursor struggles to keep up with Composer’s chat output, hangs, then shows it. Overall, I use GPT as prompt matching, then Opus 4.5 and Composer 2, depending on the task. 4.6 is fast for sure, but not worth the cost of it. Cursor needs to stop pausing 4.5, interesting that 4.6 has never gone down for me, but 4.5 does for hours, which did force me into Composer 2 - lol! worked out well.
I feel like i’m being robbed of my usage quotas / tokens with Composer 2.
It will generate me some code and throw in a bunch of useless garbage comments. I then have to give it another prompt to remove all of it’s comments.
This happens multiple times a day despite having rules / agents.md / telling it earlier in the chat.
How many tokens have I wasted doing this?
Composer 2 in Cursor — negative experience
It misses the real task, patches without thinking about the project, and often breaks more than it fixes. It says it will run tests, then doesn’t. Ask for a change and you get an unverified first guess, not alignment with what you meant — and it often feels disconnected from the last message.
For me it costs more time than it saves; hard to call it usable.
To Cursor and the model teams: custom models are fine in principle — stop overselling them in announcements and stop misleading users about what they actually get.
Translated and cleaned up with Composer 2 
.