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.
Honestly the data itself is probably worth heaps even in the company itself collapses. Early attention engineering when humans were still in the loop!!!
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.
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 lab.
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.
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.
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.
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.
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
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.
> 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...
Isn't a large user base and the data collected from those users a moat of sorts?
A moat is when you have something other's can't easily get.
Every MAG 7 / FAANG company already has more users and more data...
That's not a moat.
That's traction.
Honestly the data itself is probably worth heaps even in the company itself collapses. Early attention engineering when humans were still in the loop!!!
And its still just a vscode fork
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.
I feel like they've been targeting enterprise pretty hard. I know my company uses them, and the companies that hire us also use Cursor.
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.
I doubt it's a brand new model. It's likely just Kimi K2.5 further trained on coding.
They didn't say it's a new model... in fact they said exactly what you just said.
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 lab.
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.
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.
Well is that a statement about the quality of Opus 4.7 or about compose 2.5? :P
Full details https://cursor.com/blog/composer-2-5
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.
Check which model you're using.
The fast version of composer is the default now (which costs ~x3 as much).
My cursor costs sky rocketed recently too
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
It's always great that more companies are throwing their hat in the ring, especially focusing on value (latency + intelligence + cost)
I don't know why their model isn't on Openrouter yet. They must not have enough capacity to offer it.
It's a bit odd that they're not comparing it against Sonnet
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?
The tweet specifies that the new model is geared towards long-running tasks, which is what you'd use a model like Opus for anyway.
Did they just upgrade Kimi 2.5 to 2.6?
still uses 2.5
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 $20 plan lacking for their needs, however for my needs it fits perfectly.
The limits are probably even higher than that, i seem to get about 100$+ of usage on composer and about 45-50 usd on non composer models
Non-x link: https://cursor.com/blog/composer-2-5 (https://news.ycombinator.com/item?id=48182126)