Does compose 2 still form part of the “Auto + Composer“ or will it now be under the API allowance?
Personally I don’t find it as useful as Composer 1.5, it overthinks the assignment. I don’t need it to think for itself I just need it to do what its told
I used it for a few hours today but I found it too vague and way too verbose. I won’t be wasting tokens on it again.
Not sure if people are not prompting the models correctly, but so far Composer 2 seems very solid. I have a very complex application I am building and to make sure it does not mess up my main project I created a branch specifically just for testing composer. I never-ever use the lower tier or build in models like auto or composer but a few days ago I ran out of API calls and was forced to use composer 1.5 and I realized these models have come a long way. Composer 2 is even better and definitely an upgrade to composer 1.5 which already was solid. I honestly think these models are just making us super lazy, so our prompts deteriorate more each day as the models understand our gibberish better 
If you just think of, where we came from and what we have now… I could get some things done in GPT 3.5 and comparing those models to the newer models are like comparing an infant with a master’s degree in computer science and engineering. And it seems to still not being good enough.
Composer 2.0 is a good model, and it works well, we should really just apply some self-discipline and check our prompts and check our code afterwards. If you need a “fix all mha bugz please” AI model, then I guess we all lost the plot and should just stop programming all together. Was AI models not initially designed to be a coding assistant and now that we have these super smart models we went from coding assistance to complete coding control and moan when things go wrong.
I am genuinely pleasantly surprised by these models and are grateful I have something to work with when my API calls run out; because it has made me super lazy and I let the AI do almost everything now. tsk tsk tsk on me then I suppose.
Thanks Cursor Team, appreciate the tool allot, gave me allot of freedom while I sit and wait for the models to finish 70% of my work 
, And it seems to be really affordable too.
And before you say AI Slop, just remember that before AI even existed; multi-million-dollar companies shipped ■■■■■■ code all the time, so nothing is perfect 
For me it seems like i need to be a lot more specific with my prompt to get the right work done.
My feelings with composer 2 are mixed, recently been having issues with it failing at reading files or generating commands to search things a-lot. One of the things that bothers me with composer 2 is its too forward thinking and starts doing things I never asked it to do like update read-me files. It also thinks extremely fast even when I tried to switch down to composer 2 normal speed mode (which you removed for some reason??) It thinks way too much probably stuffing the hell out of context, and it goes so fast i cant read it, It’s frustrating I can use a models train of thought to see if its diverging from task (which composer likes to do)
I did a 40M token prompt, that took about 15min for something I have been holding out on my application for a while. I had many inline CSS and JavaScript in my razor view pages which I wanted separated. I did a test to see how well it will do just for the sake of testing a ridiculous prompt like this specifically.
It did exceptionally well that after some more testing I am actually keeping the changes… It did not drift, staid on topic and also suggested more improvements.
I actually expected it to fail; I just wanted to see where it would and how well it will do. Not only did it pass the test it did not produce one error. I am going to test some more but it genuinely looks solid.
—
All plan todos g1–g8 are marked completed in the todo list. If you want a strict “zero style= outside emails + documented exceptions” repo-wide pass, say so and we can target the remaining files systematically.
—
I’ve tested an end to end refactor on a Convex codebase, I’m comparing GPT 5.4 with Composer 2, here are my observations so far:
Token usage:
Bugbot complains:
GPT 3
Composer 4
General output quality (subjective, my own opinion):
Faster and less token usage, cheaper as well but has a tendency to over abstract, have to watch it more and ignores some basic repo level patterns, I will personally go out of my way to avoid the fast version just to save on auto quota.
On the other hand GPT models tend to be more thorough, and try and cover more corner cases, very good for legacy projects that needs lots of love & care all over the place.
Any chance we might get an API for the Composer models? These kinds of “Opus but cheaper” models are useful for a lot of other things outside of Cursor, e.g. OpenClaw.
Only issue I have found so far is that it keeps on defaulting to Composer 2 Fast, when explicitly selecting composer 2 Thinking. It’s as if it’s prioritizing Composer 2 Fast for everything.
Are Fast and normal variants the same model (with good reasoning) but just the hardware or something is different to speed up TPS? or is the fast variant a smaller model with lesser thinking budget?
If the model docs are clearer on this, it will be good.
After a day with Composer 2 I can say it is much better than Composer 1.5 so far (IMHO it was not that hard to achieve, 1.5 was pretty underwhelming for me). I use “fast” option and I like the balance it has, quality is pretty solid (but still doubt those benchmarks that compares it with Opus because as I see it Opus is still more capable model, but slower) and the speed is pretty nice (1.5 was pretty slow so it made me seek alternatives for speed).
Regarding the pricing, I seen someone asking about which “token pool” does it use, it is verified that it use from “Auto + Composer” which is logical according to the name.
So far I am pretty satisfied with it as one of my two models of choice for doing work, but despite benchmarks I doubt it will be that soon better from Opus
Is Composer 2 actually a fine-tune variant of Kimi K2.5? Users have reportedly uncovered evidence of this via the model name in the base OpenAI URL and posted about it on X. See here: https://x.com/fynnso/status/2034706304875602030?s=46
Can you confirm/deny? If the answer is “yes,” why didn’t Cursor/Anysphere note this in the release blog and documentation and include the Kimi K2.5 naming as required in its modified MIT License? “Our only modification part is that, if the Software (or any derivative works
thereof) is used for any of your commercial products or services that have
more than 100 million monthly active users, or more than 20 million US dollars
(or equivalent in other currencies) in monthly revenue, you shall prominently
display “Kimi K2.5” on the user interface of such product or service.”
I have used it mostly for code reviews and plan review and I think it works really well compared to Composer 1 (never tried 1.5).
Yup I find this as well, you have to be EXTREMELY SPECIFIC in order to actually get the UI you want with Composer 2. I am okay with that though as it does a superb job once it has guidance.
Do you genuinely believe that Cursor is going to openly admit to breaking license terms lol.
I think based on replies on the referred tweet, they may have had some deal with moonshot, coz even the moonshot guy initially responded saying they aren’t aware of any specific license but then he later deleted that tweet, so maybe they had some talk generally.
I am writing to express my profound disappointment with the recent release of Composer 2 and the “Glass” update. As a high-volume user who maintains multiple Ultra-tier plans and relies on Cursor for complex, large-scale development, the recent technical oversights and model performance issues have caused significant disruption to my workflow.
1. Inaccurate Performance Marketing
There is a stark disconnect between the marketed capabilities of the Composer models and their actual utility. While the execution speed is impressive, the logical depth is remarkably superficial.
-
Performance Gap: In my experience, Composer 1.5 and 2.0 prioritize “fast code” over “correct code.” When compared to frontier models like GPT 5.3 Codex or Opus 4.6, Composer consistently fails to resolve complex logic, often requiring manual reverts to older models to achieve a functional result.
-
Valuation: Positioning these models at a premium price point (exceeding other frontier models, comparing Composer 1.5 and GPT 5.3 Codex, feels like a misrepresentation of their actual reasoning capabilities. At present, they behave more like high-speed assistive tools than robust development models.
2. Critical Installation Oversight and Data Loss
The transition to the “Glass” version via a per-user installation path—without accounting for existing system-wide installations—is a catastrophic oversight.
-
Data Integrity: This path conflict effectively “wiped” my visibility into over 1,100 chat conversations and a 10.7 GB global database.
-
Recovery Burden: I was forced to spend several hours writing and debugging custom PowerShell scripts to strip global-only fields and manually re-format entries into the workspace DB just to recover some partial history.
-
Operational Risk: Relying on the assumption that all users utilize a per-user installation path is a dangerous oversight for a professional-grade IDE.
3. Service Plan Transitions
The issues mentioned here happened in my personal plans and personal projects. But at my work, just 5 days back, the sudden shift of frontier models to “MAX only” usage plans in enterprise plans feels like a forced migration toward the Composer ecosystem. Given that Composer 2.0 currently lacks the technical depth of the models it is intended to replace, this transition feels premature and punitive for power users.
Conclusion
I have invested significant time and financial resources into the Cursor ecosystem. However, the recent prioritize-speed-at-all-costs approach—at the expense of data stability and logical accuracy—has severely undermined my trust in the platform. I expected far more stability and technical oversight from a team catering to professional engineers.
*Request: In Plan mode, the AI must only produce plans and analysis—no file edits, no terminal commands that change state, even if the user says things like「OK」/「go ahead」. Implementation (edits, compile, git) should require Agent mode and an explicit request to execute.
*
I have tried Composer 2 in “Plan mode" but it MODIFY my code by just saying “OK”, that’s crazy and unusable which I have to explicitly tell Composer not to modify codes on Plan Mode.
I said a while ago that i was incredibly excited for any future Composer 2 release, and I have to say, I’m once again very impressed. The one thing i see is that the model requires more careful prompting to engage in deeper reasoning, but other than that, for general development purposes, it’s beyond incredible. Well done team.
Sincerely,
A Grateful College SWE Intern