> Composer 2.5 is built on the same open-source checkpoint as Composer 2, Moonshot's Kimi K2.5.

Really nice to see they're giving credit to the company and I am optimistic Kimi K open models soon will outperform Opus models

I kind of want to try it, to see if and how far they can take an open model and improve it but I really don’t miss the Cursor user experience. Constant UI changes, half-baked features, smaller and smaller limits, useless AI change attribution; I think I’ll wait for others to report if it’s any good.

The model is (like Composer 2) based on Kimi K2.5 and they claim SOTA performance for 1/10th of the cost. The tweet also mentions that they've started a new model from scratch on Colossus 2 (xAI/SpaceX Cluster). Really impressive how they've made this jump from being called the vscode fork with no moat just a couple of months ago.

I have to say the new model is quite good at the basics, I've been handing over more and more tasks from Linear straight to it instead of the copy-paste into Claude dance lately.

At this point, more of my complaints are on the harness side, which is odd since originally they were by far the best harness out there.

Support - This is pretty much non-existant, it's community support or sales support.

Interacting with GitHub - this should work and be awesome, Claude code does this well (responding to lint errors and comments). Cursor you have to poke the agent to look at the comments or lint errors, and even then it's about 10% good. Even GitHub Copilot is better here.

Bugbot - I have it setup to trigger manually, but it still seems to wake up and burn 80-120k tokens just to notice it's configured to be manually invoked. When it does run, it tells me there's no issues (but claude or copilot both find real things)

App - When you have both agent window and the ide windows, it's hard to open up the code in the right directory. A simple "cursor ." from the terminal used to do it, now it'll often open the agent window, you have to try a few times for it to work.

I love that they are running super fast, it's just hard when many of the basics break or don't work.

If these benches from their site hold up (they likely wont)

Wouldn't this compress ai revenue like 15x quickly

If they really have a 4.7 opus high equivalent at 1/16 the cost wouldn't this significantly effect all the current capex and planing

Maybe they are getting elon to cover cost

Any reason why they indexed on Kimi K2.5 model? I have tried many open-source ones in Opencode, and, in my experience (standard backend development, Java, Python, Spring, etc) Qwen3.6 is SO MUCH BETTER that's shocking. Kimi can't even get most tool calling arguments right.

  Together with SpaceXAI, we're training a significantly larger model from scratch, using 10x more total compute.
  With Colossus 2's million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.

It's very confusing that they use the same name as the very well known PHP package manager, composer

https://getcomposer.org/

They set themselves up for flack when they use whatever these evals are… they did the same for composer 2 which was evaled in close competition with frontier models, spoiler alert, it wasn’t even close in practice.

So now 2.5 is supposed to compete with opus 4.7? Sure…

I've been using Claude Code as my daily driver on a React Native + iOS codebase for the last few months. The thing that surprised me wasn't quality differences on individual edits — those are pretty close once you control for harness wiring — but how differently I'd ended up structuring my workflow around each style of tool.

Tab completion + chat-in-sidebar feels like an extension of my editing. An agentic harness feels more like delegating a 20-minute task and coming back to review. Different cognitive load, different bug profile. The "which is better" framing tends to skip over the fact that they reward different working styles.

Two things I'd watch on Composer 2.5 specifically:

1. How it handles long-running multi-file refactors that touch 10+ files. My experience with smaller models in that slot is they lose track of which files they've already edited around 30% of the way through. Frontier models keep the plan coherent for longer.

2. How it deals with non-obvious file boundaries. The thing that takes me out of "let it work" mode is the model deciding it needs to edit a config file I didn't think of. Usually that's right, but occasionally it's spelunking somewhere I don't want it to be.

The Kimi K2.5 base is interesting on its own. Open weights below frontier closed models is the thing worth watching from the harness side. If anyone's set up to fine-tune for a specific harness, this is the moment.

Ok this might be weird but I've moved everyone in my 4 person team to our team plan and costs seem to have sky rocketed compared to the individual plans. Where before most people spent 20-100 USD, now the total bill is more like 1k USD. I haven't gone into the details but it feels like I'm being scammed.

I love Cursor as a tool, but I'm skeptical bc:

1/ CursorBench is so opaque [1] that it makes it hard to trust. Not to mention the v3.1 eval is a newer iteration and there's no insight into the tasks or if the model was just tuned to max it out. Composer 2 previously scored between 60-65% on the previous benchmark eval [2] but scores between 50-55% on CB v3.1[3].

2/ I've experienced Composer 2's performance and it leaves much to be desired as a daily driver for a knowledge worker. but KWs are obviously not the target users and I can see how it's cost-efficient for executing on clearly-defined, discrete coding tasks. Obviously that's their value proposition and they're figuring out how to communicate it well to the target customer. It just doesn't feel like CursorBench is that.

[1] https://cursor.com/blog/cursorbench#building-cursorbench

[2] https://cursor.com/blog/composer-2-technical-report#performa...

[3] https://cursor.com/blog/composer-2-5

Full details https://cursor.com/blog/composer-2-5

Kudos to the team. Please consider making the model available via API!

I'm currently using Claude Code, but should I cancel it at the next renewal and switch to Composer 2.5?

Surprised this got pushed off the front page so quickly! It’s exciting to see what the Cursor team has been able to do with significantly fewer resources than the frontier labs.

I do wish they weren’t joining xAI. Something tells me there will be a contingent of researchers that departs Cursor if that merger is consummated.

It's a bit confusing to me why they'd make this 'fast' version the default, as it appears to be much more expensive than Composer 2. Wasn't it supposed to be a very cheap alternative to SOTA models?

Non-x link: https://cursor.com/blog/composer-2-5 (https://news.ycombinator.com/item?id=48182126)

The cost claim is the easy part to sell. The real test is whether it stays useful in ugly codebases, long files, and repos with a bunch of half-broken conventions. That’s where these assistants usually fall apart, even when the benchmark numbers look great.

Tested and it's good. Fast version is bad though. I like planning model in Cursor that it works more like human written design doc instead of too detailed AI plan. Seems like this is more responsible for results that model but still on fast it failed but on normal got good results.

Benchmarks measure turn-level capabilities: you feed a task into the system and then grade the result. Capability for production-level usage concerns session-level decision making: does the agent know when to stop editing, retain the right amount of context, or go back and reread the file if the state has changed?

This is not a property of the model, but a property of the discipline; it can be operationalized by what you have documented before the session begins. Without "stop editing where you can no longer follow your changes to the spec" and "go back and read the migration file before changing the schema," there is nothing to halt the process until it fails integration.

Those teams who get consistent results independent of the model being used typically do so because they have operationalized their discipline first. Those switching out models monthly tend to expect the model to supply them.

I found composer 2 pretty good as a subagent delegating tasks like auditing for bugs after finishing implementation, but hopefully composer 2.5 will be more reliable so it can be used to implement and execute long running tasks.

Say what you want about Cursor but they don’t lack for ambition.

Forking VS Code, going big on bleeding edge features like cloud agents, and now they’ve thrown down the gauntlet directly challenging frontier labs by training their own model (“much larger” than Kimi 2.5’s 1T parameters) from scratch.

They’ve been highly successful so far. Raised $50B, $2B in revenue, forecast to end 2026 above $6B. But even at these heights, they’re just not in the same league as OpenAI/Anthropic/Google.

And if building a state of the art multitrillion parameter model is not challenging enough, it’s a mountain you don’t climb just once. Every few months you need to push it farther with a new release. Fall off for a couple cycles and like Facebook you may never catch up again.

Not for the faint of heart.

I want to like composer, but I just can't.

- Its communication style is completely opposite to Anthropic models. It's not as bad as OpenAI's models, which are obsessed with "shapes", "wrinkles", hyphenated-words, and other cryptic formulations that make you feel like you're not on planet earth after a while talking to them. But it is nonetheless markedly "rude", "dry", "cold", gives off this "entitled I'm right, you're wrong" attitude. I once had composer2-fast accidentally run `rm -rf $HOME` (no harm done) as part of a bug in an install script it wrote and all it could say once it realized it was: "Running script with proper hardening". Qwen's models have clearly been distilled from Anthropic models because they have a much closer communication style and that's why I hope cursor will one day release a new family of composer models derived from that. A damn joy to use.

- It's just dumb. I don't know what they're doing with benchmarks, but for my work (python, bash, docker, whatever), cursor is just incredibly dumb. Always does in 10 lines what could be done in one. Doesn't know loads of internals of things that other models know. Never places things in the right files, constantly makes terrible edits (inline imports, edits without testing). Everything is so complicated when done by composer2, it's just a joke to me at this point. It clearly needs more handholding than Opus 4.x or GPT-5.x. I tried 2.5-fast and it seemed more of the same. And this would sort of be acceptable if it owned up to its incompetence, but it is so confidently incompetent that it's revolting.

I know that for many people the "tone" of the models is not relevant, or maybe they even prefer models like these. I simply cannot work like that.

Ever since Gemini started blowing benchmarks out of the water while being a clearly inferior model incapable of producing anything (and pretty much just doing tool calls without any feedback to the user), I gave up on benchmarks. Composer has been more of the same in that regard.

As a GPT model would say:

   "Small wrinkle: the production-ready benchmark results were tainted by real-world data points. I've assimilated the inconsistencies and added guardrails so that v2 has the right shape for future evaluations."

Congratulations on the launch! I'm interested in trying Cursor but it's very confusing what I should buy. What does the Pro $20 plan get me in usage if I only use Composer 2.5? How fast is the model?

> Really impressive how they've made this jump from being called the vscode fork with no moat just a couple of months ago.

Impressive, yes. But they still don't have a moat...

How much the RL they are doing really improves Kimi K2.5 is to be seen. So, right now, the ground truth is that they combined what they had with a strong open weights model. The RL improvement may be both marginal (since may folks report strong results with vanilla K2.6) and may mostly bias the model towards coding tasks: when a model like this is trained to be generalist, there is a tension between being good at one thing and the other, in terms of SFT and RL. You can see this in the DeepSeek v4 Flash training report for instance but it is a known fact. So if you have the GPUs and a decent RL pipeline that does not run the model you can indeed specialize it a bit more for a given task at the expenses of tasks people will not do inside Cursor. But, so far, the measurable reality is that Cursor uses an open weight model like most could do, and the RL story could be partilly a marketing move to call to Composer 2.5 more than a real strong gain, given that there is no way to verify and K2.5 was already strong. And we also know that they had to partner to do the training, which is also not a good news.

They are still a vscode fork with no moat? Like they lost about 70% of users in half a year which goes to show how there is not even the tiniest of moat.

Since the frontier is only 8-month ahead of DeepSeek, it is hard to see how model training can be a moat as all the tricks are available from open labs in China. You really just need <100m to bootstrap at this point.

This was the only way forward.

In my opinion cursor actually has one of the best harnesses again at the moment.

One would hope the vscode fork with a $50B valuation and no moat, would wisely spend the money they raised to build a moat.

It's still a VsCode fork just now with a Kimi fine tune and still no moat...

I won't debate that it turns out none of this mattered when it came to being as successful company though and kinda makes anyone who tried to roll their own instead of fork look a little silly.

>Really impressive how they've made this jump from being called the vscode fork with no moat just a couple of months ago.

With so much money and computing from SpaceX, is not so impressive.

why is that part impressive specifically? they got purchased by SpaceX, they have access to infinite compute and cash now.

& now they're still losing all of their users to Claude Code and Codex.

I doubt it's a brand new model. It's likely just Kimi K2.5 further trained on coding.

It's worth being specific:

"Will this decrease Revenue?" -- only if demand for high quality tokens is inelastic. If demand is instead elastic (grows with cheaper pricing) then revenue will likely increase.

"Will this lower earnings?" -- they have a current inference margin for their old models, and with the Elon deal in place, they have a new inference margin. It might be better or worse than their old one. If it's worse, then they'd need to see a concomitant increase in usage. If they don't, then yes it might lower earnings.

"Will this lower corporate value?" -- no - not least because this company is going to be owned by SpaceX approximately 90 days after IPO -- so all the new owner will care about is being benchmark competitive with Anthropic and oAI for the first n quarters. If they can do that, it will massively increase the corporate value of SX; it's hard to build a frontier lab.

The way I have read their benchmark results is that they trained a model to work insanely well in their coding workflow. It’s not a general purpose model.

One of the surprisingly hardest problems to solve is to get a model to use the tools you give it access to.

this thing is so awesome on fast mode, so far i am impressed, some of its observations feel similar to opus.

i use gpt 5.5 and opus 4.7 a lot every day, if i can get good results at this speed, hopefully the usage level holds up on my team plan haha

> compress ai revenue like 15x

that roughly just puts it on par with OpenAI and Anthropic subscriptions in terms of pricing per token

AI revenue has been going up while the cost per token has been rapidly falling. The Jevons paradox applies here. The cheaper software is, the more software is written. There is not a finite demand for software.

The problem with this is that we do not know the actual cost. For all we know they might be pulling an Anthropic. Subsidizing costs to get users, then increasing them later on.

I'm not sure that to be the case, it seems like bringing capabilities up and costs down merely serves to induce more demand.

Why is this comment upvoted?

It is most likely AI generated with a nice "Raised $50B" hallucination and filled with cliches ("thrown down the gauntlet", "mountain you don’t climb just once", "not for the faint of heart").

EDIT: As others have pointed out, the comment above contains hallucinations (Like the $50 billion number) and a lot of AI tells. The account doesn’t have a history of AI-like comments but the hallucinations and structure in this one are suspicious. If anything, don’t trust the numbers it cites because they’re made up.

Cursor is a team that I want to see succeed. They have stacked their company with very smart people and they’re going hard at a highly competitive market. We all win when there is more competition and more innovation.

My problem is that every few months I look at Cursor’s product offerings and maybe retry it, but it never feels like something I want to use. Part is personal preference, the other part is the fact that my combination of other tools and services just does a better job. Their biggest advantage felt like first-mover advantage when they came out early and captured market share, but at in person meetups I hear stories about companies switching away from Cursor or trying to convince their management to let them switch away. They need to come up with a compelling advantage fast, which is a hard thing to do against the other companies with their virtually unlimited budgets by comparison.

Yeah I want them to do well. I find Cursor to be a much better tool for actually working with the code the agent writes than whatever the big vendors provide.

> now they’ve thrown down the gauntlet directly challenging frontier labs by training their own model (“much larger” than Kimi 2.5’s 1T parameters) from scratch.

To clarify, the model Composer 2.5 announced in this post is not that; it uses Kimi 2.5 as a strong starting point. This is not to discount Cursor's work or future ambitions, but one of the most striking things about the last 6 months is that multiple open-source models/labs are now within striking distance of the frontier closed-sourced labs.

See eg Kimi 2.6 benchmarks: https://www.kimi.com/blog/kimi-k2-6

They have no choice but to train their own model to try and survive. They're paying API pricing for the top tier models but competing against subsidized subscriptions.

Them raising this much money doesn't mean they're successful, it only means they know how to fool the investors well. A project that is basically an extension to VSCode only adding a chat interface, isn't really worth this much money. Obviously, it's the users, but people think it's something genius and revolutionary, but no.

As a heavy user, I don't think the model is their product. Cursor is primarily a harness and lately, a specialized agent dashboard.

Composer, their in house model, is dispatched by other models like Claude Opus for individual items on a task list. No one is suggesting you write your main prompt to Composer 2.

they aren't "throwing down the gauntlet", they're trying to find ways to eke margin out of their product by owning a commodity-level coding model. it's an impressive engineering task but it's not particularly ambitious.

AI comment... BOO!