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.
It’s a near constant regression in my workflows. “Multiple agents” got destroyed recently, and the new interface for it some sort of command isn’t as good or reliable. Then you’ve got modals everywhere[1] and truncated bits (like long branch names) that make it insanely frustrating to use.
They’re constantly changing the UI without actually improving it at all. I’ll likely cancel it and use opencode for personal stuff with Deepseek and only use it at work because I have to. There was a time when I appreciated the harness but it’s becoming less useful, or at least noticeable, over time… all the while the actual UI becomes substantially more painful and awkward to use (like @ in the “agents” window being completely unable to find a file because it’s some sort of “global” scope).
One thing that surprises me about this whole segment is that JetBrains haven’t eaten these folks lunch. Their IDEs are leagues better than VSCode but their AI integration is awful by comparison (and the bar is low). I can’t even see how much of the context window I have left.
[1] it’s insane I have to answer questions in a tiny input box I cannot resize or adjust the size of. Let alone the fact the text area I input prompts into cannot be resized. Truly feels like the UI/UX is done by people without any experience.
One of the things I've came to appreciate about the cli tools like Codex or Claude is that the interface is so limited that every feature they release is still limited and constrained to the same UX limitations, whereas those "funkier" IDEs change from month to month giving me further fatigue.
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.
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.
> AI revenue has been going up while the cost per token has been rapidly falling
Every model release now has been straight price increases since what GPT 4 ? When was the last time a new flagship model decreased prices compared to the previous one ?
Not seeing that either, tried really using Opus 4.7 today, and it ended up at $50 for the same kida thing that came out to $25 last week with Opus 4.6.
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.
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 am not sure we should dismiss what they have today. Nobody has yet to come close with a full package ide that works well for coding. Is that not a moat? It is easy for my to in my head discount it, thinking that I could build something myself but between autocomplete and their workflow for agent use, it feels like they have some tangible moat emerging.
I fear the day that large parts of perfectly valid English language and punctuation are off limits for humans to use because LLMs use them too (having learned them from humans), and somebody will always whine and post low effort "slop" comments that are much more annoying and less useful than the slop itself, or even incorrectly whine about human written text that happens to match your hyper-sensitive slop detector.
Plus you are always running the risk of being rude and insulting when incorrectly labeling text actually written by humans as slop — making a jackass of yourself — and opening yourself up to being trolled by humans purposefully inserting em-dashes and catch phrases just to trigger you. That's not clever. That's gullible.
How much cognitive and physical effort and time do you put into trying to figure out if everything you read is slop, then complaining about it? If that's your job or calling in life, you could be easily replaced with AI. Find something more creative to do with your time.
If you really object to low effort slop, and not just relish it as an opportunity to whine, then how about instead of posting low effort whines about slop, you put in the actual effort to do something about it, and rewrite the slop in a way that won't trigger your slop detector, then post that instead, to train AI not to write slop.
Is your problem that it's slop, or that it's AI generated? Because your whining about low effort AI generated slop without contributing to the conversation or addressing the point of the comment you're replying to is just low effort human generated slop.
Please don't post slop while complaining about slop.
Honestly the data itself is probably worth heaps even in the company itself collapses. Early attention engineering when humans were still in the loop!!!
> Early attention engineering when humans were still in the loop
Exactly. Cursor was the first product used by tons of devs on real codebases. Just the signal "acceptance rate" is huge and can't be easily captured w/ synthetic data.
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.
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.
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.
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.
How I see this is that its so important to bundle the model with the right tooling.
Like a racecar, having the best engine doesn't help if the rest of the car lacks other winning properties (reliability, aerodynics etc).
So for Cursor, which IMO, they put themself in a strong position by having both a solid IDE __and__ a solid+cost efficient model. Those two working great in combination for the task they are designed to solve (coding) is more important than benchmarks
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…
they say it themselves in the post - behavior dimensions "not well captured by existing benchmarks". that was the exact problem with composer 2. not dumber on individual tasks, just bad at session-level decisions like when to stop editing, how much context to carry forward, when to re-read a file vs assume. you don't catch any of that in an isolated eval.
As I have said before in prior composer threads. The proof is in the usage. I am inclined to somewhat believe the results as I use composer and also take the results for the given context. It’s not a general purpose sota model. It’s a model that runs inexpensively in their coding workflow that is creating results similar to opus or gpt.
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?
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.
There's nothing to mess up. The license is MIT w/ attribution, and the attribution clause can be easily sidestepped w/o any legal repercussions. The "drama" was simply content creators going nuts over some misunderstandings and poor comms from some kimi related devs.
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.
this feels super bullish on cursor/spacexai's ability to train a frontier level model. could be truly SOTA on coding given that their RL data is this powerful
I don't think so. They're comparing it to the highest tier available models from Anthropic and OpenAI. Generally speaking, Opus is better than Sonnet in almost every way, so why have the redundancy?
Their previous Composer was already marketed as a cheap model capable of competing with SOTA on most tasks. The evals they shared back then backed this up but in my day-to-day usage it fell short across the board. Canceled my cursor subscription and switched to Claude Code a few weeks ago. It has its own shortcomings but in terms of model capability and UX quality Cursor will have a hard time competing in the long term. Elon Musk will be a very good way out for them.
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?
I use $20 plan on daily basis for more than a year now, and have yet to exhaust that limit. The plan includes $20 in api costs for non-Cursor premium models and $20 for Composer and Auto models provided by Cursor themselves.
That said, I am pretty old-fashioned coder and use LLM mostly to overcome the blank page problem, which means I review and often rewrite LLM output by hand and avoid prompt loops for a single task.
People who are aiming to not read code any more might find this $20 plan lacking for their needs, however for my needs it fits perfectly.
Really nice to see they're giving credit to the company and I am optimistic Kimi K open models soon will outperform Opus models
https://cursor.com/cli
It’s a near constant regression in my workflows. “Multiple agents” got destroyed recently, and the new interface for it some sort of command isn’t as good or reliable. Then you’ve got modals everywhere[1] and truncated bits (like long branch names) that make it insanely frustrating to use.
They’re constantly changing the UI without actually improving it at all. I’ll likely cancel it and use opencode for personal stuff with Deepseek and only use it at work because I have to. There was a time when I appreciated the harness but it’s becoming less useful, or at least noticeable, over time… all the while the actual UI becomes substantially more painful and awkward to use (like @ in the “agents” window being completely unable to find a file because it’s some sort of “global” scope).
One thing that surprises me about this whole segment is that JetBrains haven’t eaten these folks lunch. Their IDEs are leagues better than VSCode but their AI integration is awful by comparison (and the bar is low). I can’t even see how much of the context window I have left.
[1] it’s insane I have to answer questions in a tiny input box I cannot resize or adjust the size of. Let alone the fact the text area I input prompts into cannot be resized. Truly feels like the UI/UX is done by people without any experience.
One of the things I've came to appreciate about the cli tools like Codex or Claude is that the interface is so limited that every feature they release is still limited and constrained to the same UX limitations, whereas those "funkier" IDEs change from month to month giving me further fatigue.
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
Every model release now has been straight price increases since what GPT 4 ? When was the last time a new flagship model decreased prices compared to the previous one ?
One of the surprisingly hardest problems to solve is to get a model to use the tools you give it access to.
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
that roughly just puts it on par with OpenAI and Anthropic subscriptions in terms of pricing per token
Impressive, yes. But they still don't have a moat...
Every MAG 7 / FAANG company already has more users and more data...
That's not a moat.
That's traction.
That's Y.
What's wrong with using very short sentences like 'That's not X. That's Y.'?
Plus you are always running the risk of being rude and insulting when incorrectly labeling text actually written by humans as slop — making a jackass of yourself — and opening yourself up to being trolled by humans purposefully inserting em-dashes and catch phrases just to trigger you. That's not clever. That's gullible.
How much cognitive and physical effort and time do you put into trying to figure out if everything you read is slop, then complaining about it? If that's your job or calling in life, you could be easily replaced with AI. Find something more creative to do with your time.
If you really object to low effort slop, and not just relish it as an opportunity to whine, then how about instead of posting low effort whines about slop, you put in the actual effort to do something about it, and rewrite the slop in a way that won't trigger your slop detector, then post that instead, to train AI not to write slop.
Is your problem that it's slop, or that it's AI generated? Because your whining about low effort AI generated slop without contributing to the conversation or addressing the point of the comment you're replying to is just low effort human generated slop.
Please don't post slop while complaining about slop.
Exactly. Cursor was the first product used by tons of devs on real codebases. Just the signal "acceptance rate" is huge and can't be easily captured w/ synthetic data.
& now they're still losing all of their users to Claude Code and Codex.
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.
How I see this is that its so important to bundle the model with the right tooling.
Like a racecar, having the best engine doesn't help if the rest of the car lacks other winning properties (reliability, aerodynics etc).
So for Cursor, which IMO, they put themself in a strong position by having both a solid IDE __and__ a solid+cost efficient model. Those two working great in combination for the task they are designed to solve (coding) is more important than benchmarks
So now 2.5 is supposed to compete with opus 4.7? Sure…
The fast version of composer is the default now (which costs ~x3 as much).
There's nothing to mess up. The license is MIT w/ attribution, and the attribution clause can be easily sidestepped w/o any legal repercussions. The "drama" was simply content creators going nuts over some misunderstandings and poor comms from some kimi related devs.
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.
[1] https://cursor.com/blog/composer-2-5
As for the typo, s's are cheap and I've added one :)
That said, I am pretty old-fashioned coder and use LLM mostly to overcome the blank page problem, which means I review and often rewrite LLM output by hand and avoid prompt loops for a single task.
People who are aiming to not read code any more might find this $20 plan lacking for their needs, however for my needs it fits perfectly.