My experience with Composer 2 is the worst possible, and it’s making me strongly consider another LLM for my developments.
Composer 2 frequently loses context without explanation, makes gross indentation and syntax errors in Python, doesn’t properly execute commands, and performs unsolicited implementations or corrections, inexplicably undoing corrections that have already been made.
Composer 2 is simply unsustainable and has made me consume far more AI credits than necessary for increasingly simple projects.
A downgrade to version 1.5 is not only necessary, it’s urgent, because clearly, version 2 lacks the necessary capacity to meet the quality standards of its predecessors.
Ive been a heavy user of Cursor since August. In fact I have two accounts - one for work, and for personal. I have created 6 apps on a big part using Cursor. I have noticed that the new Composer2:
1-makes too many mistakes
2-missed a lot of logic or parts
3-doesnt have the ability to look at patterns in other parts of the code even if told clearly,
4-doesnt regress well, if it changes the UI, it forgets to check the routes/repo
5-even changes excellent queries to nested queries of the same thing (this blows my mind)
6-and many more
Speed is not enough measure of productivity if it doesnt do enough to scope out and set the full context of a code change. In the past one week, I have seen more simple mistakes than in the past month of work. I feel the quality of work is inferior - again depth, accuracy, completeness is more important than speed. I spend half of the time checking every part of the agents work.
I learned to love CURSOR. Even if I have CLAUDE on the side doing some of the heavy lifting, I always have CURSOR on the side to do the bulk of the tasks (CLAUDE just sucks up tokens soo fast). But with this experience, Im willing now to take time to test out another LLM (I was in windsurf at the beg for a month, tested grok/openai/gemini LLM via API in the past… maybe moving to codex and go back to pure VSCODE). With the experience I have with Composer2, it’s no longer a happy place to do work… it’s becoming a pain to spend half of the time checking, fixing, and telling CURSOR to show the gaps). My 2 subscriptions will end in a week.
The only option I can see is if I can go back Composer 1.5. Is this possible?
Also, this is an echo of a similar post below… Maybe if there are more posts like this, CURSOR will hear us. Get rid off Kimi!
Funny right after my post above, I asked Cursor to do again a similar work that I just showed/told/explained what gaps it did when updating 2 parts in B using A as pattern. And it recognized where and fixed it. But now I just asked it to do the same on C using A… The same gaps, had missing parts! Now I can attest that this is a real issue - even for pattern coding of a snippet - that Composer2 loses context so easily, and just skips easily parts is not acceptable. This could lead to a lot of code drift if left unchecked.
Real, Composer 1 was not nearly as capable. At the time though, the reason I used it for easier things was because it was so much faster than every other model, specifically GPT 5/5.1 on high, it wasn’t even on the same planet. Like using the Space Shuttle instead of a Lambo. Sure, one is nicer, but it’s still painfully slow compared to a literal rocket.
After two weeks of agony using Composer2, I too have resigned to the idea that I cant trust it to do solid code. I see first hand the kind of gaps it created on the code. Today alone, I thought I can have it do some small tasks (evaluate a bug), and it kept going in circle changing code like its playing guesswork. And the worse part, it kept giving this loongggg analysis of technical blurb (I can sense its a way to make it look like superior). But I know whats going on. So after 2+ hours I gave up (this was earlier today). So I changed the agent to Sonnet, the in a quick glance, Sonnet fixed the thing in one swoop, no blurb, scan, found, fix - very quick. The feeling of having an agent that you can feel comfortable giving work was amazing!!! The past two weeks was pure agony that the feeling after Sonnet made quick fix was a real one. Now I feel the difference between a solid LLM and a crappy one. I will not go back to Composer anytime soon.
They never intended to keep 1.5 it was a model with more parameters, that was more capable than their quantized comp2. They never intended for it to run on auto usage when they released it, but the backlash on the cost was the main reason they did that, and when they realized they weren’t fleecing us for $$$ they had to pull it down and put up something cheaper to run i stg.
Composer 1.5 was usable, Composer 2 consistently fails to listen to instructions - makes beginner coding mistakes, duplicates code all over the place. It’s OK, but unfortunately, in the time passed since Composer 1.5 came out Claude got a whole lot better and Composer 2.0 got worse.
I agree with people who say Composer 2 over-engineer stuff and make assumptions. I much prefer Kimi K2.5 which it is based on. To me Kimi is more “Claude like” in its coding style and ability to write concise yet informative documentations. Whatever additional training they did completely changed it’s behavior.
I haven’t found more mistakes in the code it writes vs. Kimi or Composer 1.5, but the formatting is less readable and I waste more time reviewing and removing the unnecessary code it keeps adding despite rules instructing against it.
Nah, Composer 1.5 was excellent at understanding what we asked for.
When a task was more difficult it assumed a smaller more doable task.
In other words, it knew what it was capable of.
Composer 2.0 thinks its Claude Opus 4.6 and acts like GPT-5.4 Mini making many really wrong undisclosed assumptions. So you can’t even predict its going to spend next 10 million tokens doing completely something else than what you asked for. That never happen with Composer 1.5, as it always asked questions instead of making stupid assumptions.
I absolutely love Composer 1.5, with its construction-level coding capability and dry, focused communication that has become my daily driver.
Enough has been said in this thread about the specific differences that I don’t care to repeat, but I can say that I won’t be switching to 2.0 and will have to choose something outside Cursor when 1.5 becomes unavailable.
It gets stuff egregiously wrong more often than not, the code it outputs is insanely complex for simple tasks, it never removes code, just adds it. It often forgets tasks & constraints literally within 1 turn.
It’s bad, so bad that it’s actually faster to write code by hand than use composer 2. Seriously, I could spend 3h writing it by hand, or I can spend 30m having composer write it and 4h fixing it and 50x the tokens.
You cannot trust it, that’s the biggest failing. You really cannot trust it with your codebase, it outputs trash, and then spreads that trash as far as possible.
IDK how you get the numbers you get on your benchmarks, but they are wrong, misleading, or fabricated.
I was making really good feature progress with auto mode and heavy guardrails, and then overnight it essentially dumped out hundreds of files of slop regardless of the guardrails, I didn’t realize this for a couple days, by then it was too late.
Turns out it was composer 2. It’s been 1 week and I’m STILL slogging through it all and trying to un-f it.
A lot of great answers to this. I have experienced with disappointment all the issues previous individuals explained myself as well.
The problem is obviously related to the fact that Cursor 2 is way smaller as a model than Cursor 1.5. The issues we all face are ones from smaller models.
One specific but critically important issue is that small models have very short attention span and are incapable of proper generalization required for important “understanding” of scope. Especially in small models we see a characteristic of transformers where they pattern-match to the opposite of what is being prompted - this is a known feature - in large models it provides proper generalization, but in small you get very easily lost in translation.
For example you say something like: “I told you that this is not the case because this and that”
The model responds something like: “I agree that this and that IS the case, you are right” - which is the exact opposite of what the user said.
Anyway, there is something way more important that we are all trying to say to you. Composer 1.5 WAS A REALLY GOOD MODEL in the sense it touched frontier level of quality! I mean personally I had a great experience with it as a user who works with all frontier models for years now every day and in highly complex tasks. It was a breakthrough and you should definitely reinstate and keep upgrading it!
Keep two models and train the small model with the larger one.
My token usage and spend has noticeably increased since the Composer 2 release. My tasks have not meaningfully increased in complexity to account for this. Seems highly profitable for Cursor to foist a model that costs them much less to run, and also costs us more to use.
