top | item 45748725

Composer: Building a fast frontier model with RL

215 points| leerob | 4 months ago |cursor.com

168 comments

order

cwyers|4 months ago

The lack of transparency here is wild. They aggregate the scores of the models they test against, which obscures the performance. They only release results on their own internal benchmark that they won't release. They talk about RL training but they don't discuss anything else about how the model was trained, including if they did their own pre-training or fine-tuned an existing model. I'm skeptical of basically everything claimed here until either they share more details or someone is able to interpedently benchmark this.

criemen|4 months ago

I understand where you're coming from, and I'd love to have learned about pre-training vs. off-the-shelf base model too. But

> their own internal benchmark that they won't release

If they'd release their internal benchmark suite, it'd make it into the training set of about every LLM, which from a strictly scientific standpoint, invalidates all conclusions drawn from that benchmark from then on. On the other hand, not releasing the benchmark means they could've hand-picked the datapoints to favor them. It's a problem that can't be resolved unfortunately.

infecto|4 months ago

Disagree. The ultimate bar which is easily measurable, do users find value in it. Benchmarks are mostly meaningless especially in my opinion where cursor shines which is the tool chain. You can go try composer yourself today and see if it’s valuable to you.

NitpickLawyer|4 months ago

Does it really matter tho? At the end of the day, what matters most is if real users find it useful or not. And cursor has that data (both historically and in real-time). Thousands of accepts/rejects >>> any benchmark that you can come up with. That should allow them to iterate on it, and make it better, eventually.

Benchmarks have become less and less useful. We have our own tests that we run whenever a new model comes out. It's a collection of trivial -> medium -> hard tasks that we've gathered, and it's much more useful to us than any published table. And it leads to more interesting finds, such as using cheaper models (5-mini, fast-code-1, etc) on some tasks vs. the big guns on other tasks.

I'm happy to see cursor iterate, as they were pretty vulnerable to the labs leaving them behind when all of them came out with coding agents. The multi-agents w/ built in git tree support is another big thing they launched recently. They can use their users as "teacher models" for multiple completions by competing models, and by proxying those calls, they get all the signals. And they can then use those signals to iterate on their own models. Cool stuff. We actually need competing products keeping eachother in check, w/ the end result being more options for us, and sometimes even cheaper usage overall.

jonasnelle|4 months ago

Cursor has the best Tab model, and I feel like their lead there has kept growing - they're doing some really cool things there. https://cursor.com/blog/tab-rl

I wonder how much the methods/systems/data transfer, if they can pull off the same with their agentic coding model that would be exciting.

enraged_camel|4 months ago

Tab model is fantastic but I wish it was somehow aware of the conversation happening in the currently active AI chat session.

srush|4 months ago

We also are big Tab users here at Cursor. In the blog we talk about the motivation for this project came from thinking about a Tab-like agent.

oersted|4 months ago

I agree, I tried to switch to Zed this week, and I prefer it in all respects, but the tab model is much worse, and it made me switch back. I never imagined I would care so much about a feature I felt was secondary.

I actually find myself using the agent mode less now, I like keeping code lean by hand and avoid technical debt. But I do use the tab completions constantly and they are fantastic now ever since they can jump around the file.

vidarh|4 months ago

I feel like that's like having a lead in producing better buggy whips.

I run Claude Code in the background near constantly for a variety of projects, with --dangerously-skip-permissions, and review progress periodically. Tabbing is only relevant when it's totally failing to make progress and I have to manually intervene, and that to me is a failure scenario that is happening less and less often.

dagss|4 months ago

It's great. BUT: Wish they had selected another shortcut like shift+tab.

Every time I write code myself I find myself racing the AI to get an indentation in before the AI is done... gets annoying

typpilol|4 months ago

What makes it and different from vscodes copilot completions?

srush|4 months ago

Hi everyone,

I am an ML researcher at Cursor, and worked on this project. Would love to hear any feedback you may have on the model, and can answer question about the blog post.

coder543|4 months ago

Impressive systems write-up. A question: if Composer is an RL finetune on an open model, why keep weights closed? The edge from a slightly better checkpoint erodes quickly in this market, it's not a durable advantage. Composer protects Cursor's margins from being squeezed by the big AI labs, but that is true whether the weights are open or closed, and I think Cursor would have more lasting benefit by generating developer goodwill than from a narrow, short-lived advantage. But, that's just my opinion. I personally find it hard to get excited about yet-another proprietary model. GPT-5 and Sonnet 4.5 are around when I need one of those, but I think the future is open.

Agingcoder|4 months ago

It's stunning.

I don't use these tools that much ( I tried and rejected Cursor a while ago, and decided not to use it ) but having played with GPT5 Codex ( as a paying customer) yesterday in regular VSCode , and having had Composer1 do the exact same things just now, it's night and day.

Composer did everything better, didn't stumble where Codex failed, and most importantly, the speed makes a huge difference. It's extremely comfortable to use, congrats.

Edit: I will therefore reconsider my previous rejection

WanderPanda|4 months ago

Why did you stop training shy of the frontier models? From the log plot it seems like you would only need ~50% more compute to reach frontier capability

chaidhat|4 months ago

Which model did you distill it from? Great work! PS getting a few scenarios where it doesn't follow rules as well as sonnet 4.5

embedding-shape|4 months ago

Do you have any graphs handy that kind of replicates the one used first in the blog post but a bit less ambiguous, maybe without model grouping? I feel like it would have been a bit more fair to include proper names, and individualize them rather than group everything together by something, and then present your own model on its own.

alyxya|4 months ago

Is the new model trained from scratch? What training data went into it?

dfltr|4 months ago

Is it true that Cheetah is Grok Code Fast 2? Does this mean that the new Cursor model is also based on Grok?

dlojudice|4 months ago

Congratulations on your work. I spent the day working with a mix of the Composer/Sonnet 4.5/Gemini 2.5 Pro models. In terms of quality, the Composer seems to perform well compared to the others. I have no complaints so far. I'm still using Claude for planning/starting a task, but the Composer performed very well in execution. What I've really enjoyed is the speed. I had already tested other fast models, but with poor quality. Composer is the first one that combines speed and quality, and the experience has been very enjoyable to work with.

juanma0216|4 months ago

I prefer the approach of focusing on faster models despite their lower intelligence because I want my IDE to fly when I can see the code. I find this useful when I need to manually debug something that any model is able to do, so I know it's going to fail but at least it will fail fast. On the other hand, if I need more intelligence I have my other CLI that doesn't allow me to see the code but gets the planning and difficult code done.

pdeva1|4 months ago

is Composer a fine tune of an existing open source base model?

smg|4 months ago

Can you please tell us more about how you used Ray for setting up the RL infrastructure?

ripped_britches|4 months ago

Amazing work! The UX is great.

GPT-5-codex does more research before tackling a task, that is the biggest weakness for me not using Composer yet.

Could you provide any color on whether ACP (from zed) will be supported?

az226|4 months ago

How many times have you needed to reset the optimizer during the RL training cycles?

carlosbaraza|4 months ago

How do you work with multiple agents?

jasonjmcghee|4 months ago

Maybe I'm an outlier but Sonnet 4.5 quality is about as low as I'm willing to go.

It's generation speed is not the problem or the time sink.

It's wrestling with it to get the right output.

---

And just to clarify as maybe I misunderstood again but people are comparing cursor to Claude Code and codex etc here- isn't this whole article all cursor just using different models?

swyx|4 months ago

> Sonnet 4.5 quality is about as low as I'm willing to go.

literally a 30 day old model and you've moved the "low" goalpost all the way there haha. funny how humans work

alyxya|4 months ago

There’s two different kinds of users, on one side people are more hands off and want the model to autonomously handle longer tasks on its own with minimal guidance, and on the other side is users who want to interactively collaborate with the model to produce desired results. Speed matters much more for the second case, where you know what you want and just want the model to implement whatever you had in mind as quick as possible. Intelligence/ability matters more for the first case when you don’t have full understanding of all the code. I think it’s context dependent for me where more serious work tends to be more interactive. The intelligence of a model doesn’t make up for issues due to lack of context to me.

srush|4 months ago

Agree that Sonnet 4.5 is an excellent model. Would be curious to hear your experience using Composer though, it's quite good.

timcobb|4 months ago

Same... I've found that using a non-Claude model just ends up being more expensive and not worth it. "Auto" tokens are hardly free, and I've had plenty of experiences putting "Auto" to work on a "simple" seeming task to have it consume like 1 USD of tokens quite quickly while producing nothing of value, when I'd replay with Claude 4.5 Sonnet non-thinking and it would provide a solid solution for 0.5 USD.

solarkraft|4 months ago

The reason I pulled out the comparison is to highlight how serious they are about all the important parts that make or break the AI coding experience - speed being very important to me. I’d rather catch my model doing the wrong thing quickly than having a higher chance of one-shotting it at the cost of having to do a lot of specification upfront.

NaomiLehman|4 months ago

gpt-5-high is as low as i can go :]

stared|4 months ago

While I am excited to see a new model, I am skeptical when there is so much vagueness - charts with "frontier models" without actually spelling out which ones, charts with no numbers (time axis, or in one chart - entirely).

srush|4 months ago

There is a footnote that should help with the models. Training is a harder thing to report on, but roughly our finding here is that RL scales.

solarkraft|4 months ago

People on here love to be contrarian about Cursor, but I’ve tried all the popular alternatives (Copilot, Claude Code, Codex, Gemini CLI, Cline) and found Cursor’s overall experience to just be unmatched. A big part of that is its speed, another its reliability.

It’s the only coding agent I’m actually really motivated to use out of the box because it really does make me feel more productive while the others keep messing up the project, from way too large changes I didn’t ask for all the way to constant syntax and request errors.

It’s the only coding agent I’ve used that feels serious about being a product rather than a prototype. Their effort in improving their stack is totally paying off.

pqdbr|4 months ago

I dropped cursor for the precise reason you mention: reliability.

Countless times my requests in the AI chat just hang there for 30+ seconds more until I can retry them.

When I decided to give Claude Code a try (I thought I didn't need it because I used Claude in Cursor) I couldn't believe how faster it was, and literally 100% reliable.

EDIT: given today's release, decided to give it a go. The Composer1 model _is_ fast, but right at the second new agent I started I got this:

> Connection failed. If the problem persists, please check your internet connection or VPN

infecto|4 months ago

I too have tried them all and have settled with Cursor being the best. That said I see the current space split between folks like me who know generally what I want built and appreciate a tool that helps me get to goal quicker and on the otherwise of the spectrum, folks who want the tool to orchestrate most of the engineering. I have no opinion on which is better but for me I sit on the first camp. In that camp Cursor is by far the best tool.

saberience|4 months ago

I used Cursor for the total of one day (paid for a year subscription), discovered Claude Code later that day and havent opened Cursor since.

Note, later I started using Codex and now Codex is my daily driver, Claude Code for problems where Codex fails (not many), and again Cursor is never used.

They were the first mover but Codex (in my opinion) blows Cursor up into 1000 tiny pieces. It's just so, so much better.

psygn89|4 months ago

Yep, it just works seamlessly. Sure, it hangs sometimes, but their UI allows you to retry or undo changes to an earlier point in the conversation easily. The autocompletion is nice as well and pretty satisfying to tab through the small and menial things when refactoring.

rtfeldman|4 months ago

> I’ve tried all the popular alternatives (Copilot, Claude Code, Codex, Gemini CLI, Cline)

Can't help but notice you haven't tried Zed!

ramon156|4 months ago

You tried Claude and still prefer cursor?

OsrsNeedsf2P|4 months ago

One thing no competitor is serious on is average response completion time. Cursor lapped everyone there

srush|4 months ago

There are lots of good models we like here. But we agree that getting the right point on the smart+fast graph can make agentic coding feel really good.

(Cursor researcher)

nu11ptr|4 months ago

I love Cursor. I've tried Copilot/Claude/etc. but keep coming back to Cursor. I just want to work, and Cursor tab complete is dang accurate, esp. for refactoring tasks.

Sammi|4 months ago

I tried going back to VS Code + Copilot a month ago. I only lasted 4 days because it was to bad. It was super slow and gave poor suggestions, but mostly it just flat out did not suggest anything. Cursor feels snappy in comparison and the suggestions are more often than not useful. The most annoying thing about Cursor tab complete, is that it is so fast that when I am doing something unusual then it will keep on jumping in with useless suggestions. They have a snooze function for this though.

neuronexmachina|4 months ago

For anyone else who was wondering, it looks like the within-Cursor model pricing for Cursor Composer is identical to gemini-2.5-pro, gpt-5, and gpt-5-codex: https://cursor.com/docs/models#model-pricing

($1.25 input, $1.25 cache write, $0.13 cache read, and $10 output per million tokens)

lubujackson|4 months ago

I'm curious if their near-term expectation is that this is be better than these models or is this a model they tend to use in Auto mode, or if the focus is really if you want speed...? I guess my question is why would I actively chose this over Auto?

SafeDusk|4 months ago

I think both Cursor and Cognition and going in the same direction of SWE-grep[0].

SWE-grep was able to hit ~700tokens/s and Cursor ~300token/s, hard to compare the precision/recall and cost effectiveness though, considering SWE-grep also adopted a "hack" of running it on Cerebras.

I'm trying to kickstart a RL-based code search project called "op-grep" here[1], still pretty early, but looking for collaborators!

[0]: https://cognition.ai/blog/swe-grep [1]: https://github.com/aperoc/op-grep

swyx|4 months ago

excited to see how far you get with opgrep!

toobulkeh|4 months ago

I used the new system tonight and it felt like a definite downgrade. Generated a few non-working basic apps, couldn’t handle CSS in a NextJS environment. Terminal context didn’t work. And it went back to not reasoning through the problem until resolution. And kept slowing down.

I’m assuming major release vs stable, but this is pretty lackluster so far. Switched back to Sonnet reasoning. Here’s to improving!

carlosbaraza|4 months ago

Could anyone explain how to use multiple agents and subagents in Cursor, Claude Code, or others? It is already challenging to me taming one model doing work, let alone synchronizing multiple parallel workers.

Do you have to split the plan in parallelizable tasks that could be worked in parallel in one codebase without breaking and confusing the other agents?

asdev|4 months ago

you can use git worktrees and just have multiple Claude Code terminal instances working on each worktree. That way they don't clash, just delete the worktree when the task is done.

koakuma-chan|4 months ago

I just gave it a try and it's reaally fast. Didn't expect this from you Cursor, good job.

kilroy123|4 months ago

What I can't stand about cursor is the constantly changing and confusing billing and usage.

I think competition in the space is a good thing, but I'm very skeptical their model will outperform Claude.

netcraft|4 months ago

I love cursor, the tab completion and agent mode. But I really dislike vscode after using intellij for so many years. I really wish the underlying editor was better, or I could get cursor features in intellij instead. The editing of the files is mostly fine, but its everything else around it that a full IDE provides thats just so much better. Right now its intellij + claude code for me, and its fine, but I wish I could get the AI power of cursor in a better package.

pbowyer|4 months ago

Intellij's tab-complete is coming along; it's hit and miss if it will work but for similar edits I'm finding it picks up the pattern quickly and I can tab - tab - tab to make them happen.

Still not up to Cursor standards though :)

Jcampuzano2|4 months ago

Building off of VSCode was probably Cursors silver bullet and the best decision they could have ever made.

It made migrating for everyone using VSCode (probably the single most popular editor) or another vscode forked editor (but at the time it was basically all VSCode) as simple as install and import settings.

I do not think Cursor would have done nearly as well as it has if it didn't. So even though it can be subpar in some areas due to VSCodes baggage, its probably staying that way for a while.

carlosbaraza|4 months ago

Cursor 2.0 keeps crashing on me while having an agent running and opening the IDE part of the application. I might have to rollback.

amilich|4 months ago

Hey - really sorry to hear this - could you email me andrew@cursor.com? Here are 3 suggestions to try- 1. Reset your settings.json - if shared with vscode, sometimes settings can cause perf regressions 2. Could you try cmd-shift-p -> "capture and send debugging data"? Will send us some profiling data to debug 3. Clear your user data (will delete chats) as a last resort - cmd-shift-p, "reveal user data," close the app, then delete this folder and restart the app

Jayakumark|4 months ago

This looks like a model RLed on top of Qwen3-Coder or GLM 4.6 as per their graph and foot note.

romanovcode|4 months ago

Where is the comparison with Sonnet 4.5? That would be the only thing that matters, really.

matheist|4 months ago

> "Best Frontier" includes GPT-5 and Sonnet 4.5, which both outperform Composer.

swyx|4 months ago

swyx|4 months ago

my very small nit is... why is the model called Composer?? of all things?? when there was already a Cursor Composer from 2024.

Cursor Cheetah wouldve been amazing. reusing the Composer name feels like the reverse OpenAI Codex move haha

asdev|4 months ago

is Cursor Bench open? Would like to see an open benchmark for agentic coding

srush|4 months ago

Unfortunately not, as we used our own internal code for the benchmark. We would also like to see more benchmarks that reflect the day-to-day agentic coding use.

timcobb|4 months ago

I wish it was easy to find out how much it costs relative to Claude :)

sebdufbeau|4 months ago

As a stealth model, it was priced as $1.25M in / $10M out

Right now, it seems free when you are a Cursor Pro user, but I'd love more clarity on how much it will cost (I can't believe it'll be unlimited usage for subscribers)

skeptrune|4 months ago

Facts. They really need to make pricing more clear across the entire product.

ciphix|4 months ago

The metrics in the post seem quite abstract. Does anyone know the detailed metrics of this mysterious model? Was it fine-tuned from open models or trained from scratch?

ianberdin|4 months ago

Feels like the comments are fighting of prepaid influencers.

alyxya|4 months ago

I wonder if this custom model is trained on cursor users. There’s a lot of potential on how much better a custom model could be the closer it is integrated with the product. Having the model learn to adapt to different user preferences would make it stand out compared to memoryless frontier models.

Sammi|4 months ago

The fact that you are wondering this is bad. You definitely should know this. _ALL_ the online ai providers are training on your data. They have more expensive enterprise plans if want to opt out.

ibash|4 months ago

Very cool, congrats!

numbers|4 months ago

Please keep the naming of your models sane, I'd like to know that composer 1 is the first model and composer 2 is second but composer 1o is not yet another 1 variant that's actually newer and better than 2, that's just dumb. Not that you're doing that, some other companies do that.

srush|4 months ago

We will do our best. Luckily I don't think there are major telecom companies called Composer-2.

arresin|4 months ago

Same price as GPT-5