CursorBench 3.1
cursor.com
cursor.com
Cursor's benchmark finds that Cursor's model (Composer 2.5) is basically as good as Opus 4.8 max and GPT-5.5 xhigh, but at a fraction of the price.
Artificial Analysis' testing shows Composer 2.5 to be pretty far behind: https://artificialanalysis.ai/agents/coding-agents. You look at the DeepSWE benchmark (which is probably the hardest to game at this point) and GPT-5.5 xhigh gets a 64, Opus 4.8 max gets 56, and Cursor 2.5 gets 16.
I don't doubt that Cursor works well for some people. It's beating DeepSeek v4 Pro in the DeepSWE benchmark and that's a very capable model. But I'm skeptical of the claims that it's a competitor for Opus 4.8 and GPT-5.5. It just seems convenient that their model does so well on their own benchmark while third party benchmarks have it far behind. Maybe it's a really great benchmark and a better measure than third party ones - I'd love for a cheap model to do as well as the expensive ones.
Cursor's Composer 2.5 is refreshingly fast by comparison, but I don't trust it as much as Opus. And Composer doesn't have 1 million context...
I mentioned my frustration with Composer in another thread and why I rely on Opus, but Opus, atleast via Cursor is practically unusable for me M-F 9-5 EST. As a result, I have modified my working schedule outside those hours when I use Opus. On weekends and nights, Opus via Cursor is at the same speed as Composer but vastly superior quality where it's not even comparable.
Composer is not 20% worse than Opus for me. Composer hands me a quickly put together college project that was started the night before it was due. Opus hands me a actual production ready deliverable that I can defend if I was sued in court.
There's also issues with cost calculation (as that harness doesn't use caches) and so on as reported on their github issues.
None of the benchmarks are perfect, but that does explain a lot of the variations between benchmarks.
for most tasks is capable and very cheap, for a days worth of tasks is costing about $10
I do feel that they've really upped their game with composer this year though.
Didn't notice yours until now.
Vim gives you that by highlighting the lines and running `:sort` for free
Ironically, their benchmark might be more accurate than artificial analysis for a narrow slice of things that Cursor's Eigencustomer is really interested in. Otherwise I'd take it as just another data point.
Overall, some benchmarks show Composer doing well, others not so much. We think the model is very capable at the given price point. There's lots to improve! If you see any specific behaviors or places the model isn't very good, lmk here or can email me lrobinson at cursor.com.
Grok build only gave me roughly 10 hours of use for $40 for the entire month...
I don't even care about long horizon, can I use it a reasonable amount of time through the month? I use AI for hobby projects, Claude gets me quite far, but I tire of dropping $100 every month. I'm not sending my money to some Chinese firm that now has access to my computer.
1) Fork VSCode and pretend that they made some kind of "AI IDE"
2) Fork a Chinese model and pretend that they trained their own model
Company with zero moat and terrible products that keeps lying to everybody
The "price point" comparison is a lie though because Composer is only available with a monthly Cursor subscription, and Cursor's external-model-per-token charges for other models are not representative of what other models' monthly subscribers get. An OpenAI $200 subscription gets you at least as much GPT 5.5 as a $200 Cursor subscription gets you Composer 2.5.
Your skepticism is well-founded IMHO. I have found that if you are one-shotting a Django/Next CRUD app, a typical React/Vue UI, shell scripts or GitHub Actions, Composer 2.5 is fantastic!
But for anything outside the median of the last decade's web development - like free-body physics, kinematics, or optimization - Composer is horribly unpredictable.
That's what makes it _dangerous_ IMHO.
It isn't universally trash! Rather, it confidently makes subtle, incorrect assumptions. It will hallucinate formulas that don't appear in your specification and design docs. Then write tests that pass it.
It inserts tiny footguns that require you to scrutinize every single token it generates. At that point, I would rather be coding by hand.
Opus 4.8 max, on the other hand, refuses to guess, atleast the way I have set it up. If there's any ambiguity about the implementation or how tests should be written, it stops and asks me for clarification. I actually trust the output without worrying about hidden disasters and ticking timebombs. I can confidently review the test suite, add a few edge cases on my own, spot check the code and be comfortable knowing there are no disastrous footguns lurking in the shadows only to come out in the darkness of production deployments.
Let me repeat - Opus 4.8 max stops and asks me for clarification. It writes the tests I would have written. It writes tests that fail, exposing gaps and errors, that then allows me to iterate.
Composer 2.5 OTOH will run with whatever it decides I meant and write something that steals productivity, not add to it.
Same harness (Cursor), same rules, same prompts, vastly different outcomes!
Yes, Opus is far more expensive, but it's worth it for the time saved on review and refactors, which are our current blockers.
The real friction is that Cursor's marketing is so aggressive that the people paying the bills look at my Opus usage and demand to know why I'm not using the cheaper alternative!
It's an impossible argument to win when the rest of the company's devs are happily building standard web apps on Composer without issue, blissfully unaware of how the model not only falls apart but is just unreliable on harder engineering problems.
Fable 5 is on a league on its own. If history in the LLM space is any predictor of the future, in ~6 months (Q1 2027) we should have open weight models that are competitive with Fable 5. Without considering what it will take to run such a thing, I would be extremely excited to have open access to such a capability. Great times ahead!
Putting that aside, I spend all day every day implementing very, very hard things right on the edge of what agents are (barely, sometimes) capable of, and I have had to keep Opus on max for things that need 'real validation' for a while now. And that has felt like 'the only way' to get Opus to perform even close to 5.5 xhigh. I'm only using Opus at all because GPT-5.5 in the subscriptions only has a small (400k, but 258k effective) context window.
The difference is that 5.5 xhigh is extremely fast in most practical cases, both efficiently implementing _overall_, and responding very quickly with great adaptive thinking if you ask it something that it doesn't have to think about. Opus 4.8 Max will needlessly chew on everything and can take hours to implement even simple things, so I can mostly only use it for planning/review.
Fable is much much better at adaptive thinking / responding quickly (although probably still worse than 5.5 xhigh), and... I think folks have said enough elsewhere about its strengths and weaknesses. Sadly still not a reliable implementor for my hard tasks though (that's still GPT's domain) – it tends to leave big, dangerous holes hiding inside implementations unless babied.
Do you find that makes a difference in your work? I've been using 5.5 high/xhigh to optimize and benchmark a C codebase, and just reading the initial code virtually fills the first context window. A session will auto-compact 5-15 times, but it seems to do okay in spite of that because the task is mainly focused on the latest window each time.
I think for programming the strength of GPT over Opus is winning here over the context window.
On this, absolutely!
I more often use Opus for planning than for implementation. In those cases I really do need the very large context window, because the agent has to read in a bunch of my code base and a bunch of previous plan files and product context and such, to understand what we're talking about.
And then I need to go back and forth with it over a really extended period: getting into a bunch of details, asking it to load how things already work so that we can discuss options for evolution of those, etc.
For that kind of thing, compaction completely destroys its effectiveness because even if you try to serialize out all the decisions made in the conversation into a plan file, the agent still loses e.g. the plan files and code files that it's read in that are adding sharp edges to its understanding of the scope of what's being planned.
For implementation or something like what you're describing in the vein of benchmarking, often I can get away with compaction. Although even then, if the agent needs to have a lot "loaded" into its head, to implement something very, very subtle, complex or far-reaching, in those cases it can be really detrimental if it compacts.
for supporting evidence, see first chart here: https://www.anthropic.com/news/claude-fable-5-mythos-5
Is a single thing in your post demonstrable, or are we just supposed to take your word for it? Because all of this stuff sounds laughably subjective.
Just check out any conversation on dynamic vs static typing, talk to a Rust zealot, or ask a backend engineer if microservices were a mistake.
It's unfortunate, and it makes it hard to have proper discussions on these subjects. It would be worthwhile to figure out how we can have more constructive arguments.
Frankly, it feels like we should just sidestep arguments entirely and just all contribute our messy data/reports, and then see how we can meld all of it together, to find the best answers for our individual situations.
Probably a good use of frontier AI, melding all of that!
On performance: just grab CC + Codex and try Opus 4.8 xhigh and GPT 5.5 xhigh side by side. Ask them a trivial question about something that's already in their context. Opus will churn for 30 seconds, and GPT 5.5 will respond in about three seconds. If you try the same with Fable 5 you'll notice way better adaptive thinking than Opus (it'll quicker than Opus, even on xhigh – although often still slower than 5.5).
I have many, many times done 'Opus xhigh, Opus max and GPT xhigh all tried to implement something' – Opus max is... hours and hours. Opus xhigh is usually ~1.5-2x GPT 5.5 xhigh. This feels like a pretty straightforward generalization of the first point. Again, just try racing three agents and see what you get.
As far as 'right on the edge of what they're able to do', my specific tasks don't matter. Just find something that no matter how hard you try, with however many agents or combinations thereof, with arbitrarily detailed plans, agents can't seem to implement without massively mistakes or a hollowing-out of 'the point' of the implementation... and then try it on the 'following generation' of models. I've been doing this repeatedly with coding agents since I turned aider into a CC-like coding agent in early 2025 (this was my second one, my first modern-style coding agent was in Jan 2025): https://github.com/Aider-AI/aider/pull/3781
A couple of examples of the latter thing that I tend to work on are database internals (indexes, query planner stuff, etc.; I built the DB in full before agents, it just works on it with me), very advanced UIs (try making a beautiful Rolex-like interactive visualization of the internals of a mechanical watch with Opus and see how far it gets – not very), and 'hardcore product questions' (all agents kinda suck at schema – Fable far less than prior ones). I have dozens and dozens of these that they can't do, though.
I don’t remember this always being true.
A brainwave: perhaps GLM or DeepSeek could be integrated into the mix for the purposes of red-teaming the code. Fable has been blinded to security by design[0], and the open models are pretty decent at it.
[0] It's not clear what the situation with GPT-5.6 will be but the blog suggests similarly over-cautious safety filters.
Amusingly the posts for recent Opus releases brag that they successfully made it worse at security! "during its [Opus 4.7] training we experimented with efforts to differentially reduce these ["cyber"] capabilities"
And yeah, it's a bit depressing to think that 5.6 might be similarly nerfed. Less secure software for us all, I guess... except BigCorps. :(
To put their own model out in front?
it's fascinating that I used these same exact words to express my distaste for Composer and my preference for Opus. I suspect, the domains and problems we are trying to solve need to be shared. I wrote about it here: https://news.ycombinator.com/item?id=48766275
Would love to reach out to discuss more, if you're ok with it, or absolutely feel free to do the same as my email's in the profile like yours!
I am not building rockets, but have been quite impressed. All the models do dumb things sometimes, it has done the work I have asked it to pretty well though and has done to me some impressive work.
It is fast on Grok, for other models I have worked extensively with I think it is better than gemini 3.1 (3.5 and antigravity for me is worse than the prior gemini cli). And is comparable to Opus 4.6. (Have not used the more recent models in Claude Code.)
Composer did a competent but not amazing job with a good plan. What I really liked though is it was fast! Opus could take 30 minutes to do something Composer would get done in 5-10 minutes. Of course the output wasn't perfect, but that's why I'd do a cleanup pass using Opus or Codex.
It's all a balance though, constantly changing and completely dependent on the problem you're solving. I just remain flexible and adapt my process to what's working best in the moment.
Have you settled on what the clean up pass should look like? Or, do you keep experimenting with it?
In case one might not have been aware: Composer 2 was Kimi Base 2.5 post-trained (RL'd) by Cursor: https://news.ycombinator.com/item?id=48507474. Composer 2.5 might be something totally different.
Doesn't matter which model wrote the code, they all make mistakes. This is the same stuff I'd do with any junior engineer's PR and it leads to better quality outcomes (something I care about and am finding hard to let go).
Fable is particularly good at having this back and forth I'm discovering.
Although, the benchies are always tricksy ... On DeepSWE, GPT-5.5 beats Opus-4.8, by a fair margin, but on FrontierCode, the situation is the other way around.
The only benchmark you can trust is your actual workload!
With that being said, Cursor Bench is nonetheless a fantastic gauge of LLM quality and performance. The most interesting outcome of Cursor Bench is Fable 5 to GPT 5.5, its almost a perfect continuous line. As of the report, Fable to GPT might be the new standard of agentic programming/building
I did my own rounds of tests for a lot of models.
I made multiple rounds of comparisons for issues in my open source project using Fable 5, GPT 5.5, Opus 4.8, and Composer 2.5
Made them work on various complexity issues
Here are the reports, i recommend using a Sonnet 5 or better model to summarize them because its quite a bit of information to consume across all the various tests: https://www.richkuo.com/#llm-battles
My takeaway from Composer 2.5 is that its better to use Opus or GPT for planning, and then Composer to build, use Opus/GPT to pr review, and have Composer fix findings, and loop that process. Incredibly token efficient, fast, and gets the same quality as if you worked the whole thing with Fable/Opus/GPT.
Which is exactly why every credible comparison now represents cost associated with completing a task, not arbitrary input and output token costs.
Not sure how much "real intelligence" is to be found in Mythos & Sol, but at this point, ignoring the intelligence gap, I find it totally impressive that the likes of GLM, Kimi, Qwen, MiMo hold their own at 2x to 4x less cost, and work for my use case just the same.
Fable is using less tokens to achive that same tasks compared to sonet and opus. If so that is a good thing. It feels like we for a while there was spitting out tokens to get a better result. If the model themselves are getting better without generating more tokens that feels like a real win.
Q1: Why is number of steps relevant in this graph? What does it tell us?
Q2: and why have they flipped the horizontal graph so that 0 is to the right and not at origo? Is that some kind of new smart thing? can't say i have seen it before
The other models however are reasonably where I’d expect them to be from experience piloting all of them. Fable is outclassing everything at most things at 10x the cost, but sometimes it isn’t a choice between cheap and expensive, but expensive and possible; I’ll need to learn where that boundary is just as it was the case with other models.
And given that you can only use Composer with a Cursor monthly subscription, cost comparisons are pointless since an equivalently priced OpenAI subscription gets you just as much usage of the better model.
Can we get a count of people that have had Claude read irrelevant documents or perform unnecessary web searches even when told not to from the beginning?
I'm starting to wonder if this increased token usage is inadvertently bleeding into how Anthropic actually trains their model, especially leading up to IPO. As older models are deprecated and users are forced onto newer models, if the default is less efficient and more token expensive that directly results in higher "profit" for Anthropic in terms of the consumption their users have to tolerate - lest they jump to a competitor.
I keep Claude around for some specific tasks:
- Linked up to Figma MCP to implement front-end stuff
- Data analysis, in the "Connect AI to a data source and ask questions" way. I've tried both Opus 4.8 high and GPT 5.5 high for this and Opus is stronger because it gets the intent in the question better
I used to keep it around for planning too, but the 4.8 plans have had more holes than swiss cheese.
Related: Sonnet 5’s new tokenizer increases token usage by 30%. (https://simonwillison.net/2026/Jun/30/claude-sonnet-5/)
Compared with how much it can do in such little time, it's still far less than even a junior engineer.
I only reach for Claud when i need to plan something big or want to have a sparring partner to fire of some ideas.
I think what a lot of people don't realize is that you don't need a fronteer model for 80% of coding tasks. Composer 2.5 is often more than good enough, less token hungry and way faster
I'm working on a fairly routine full stack web app that isn't doing anything incredible. Once I had the patterns I wanted in place, it's been very capable of following those with new work. I also don't ever give it long running tasks, it's always focused and small chunks.
My typical work flow is 1. /grill-me feature description 2. Create a plan 3. Manually review plan and tweak as needed (usually very little to none) 4. Build the plan
All with Composer 2.5. Earlier on in the project I used Claude and GPT for #1 and #2.
I find it really hard to justify the other models for the performance/cost I'm getting with Composer 2.5. Maybe it's not as strong as the frontier models, but it's been plenty good enough for my use cases.