Your AI coding assistant just got a competitor — from the very tool millions of developers already use to write code faster. Cursor, the San Francisco-based startup that popularized “vibe coding,” unveiled Composer 2 today, its first in-house AI model trained exclusively on coding data. And it’s gunning directly for Claude and GPT.
The announcement lands on March 19, 2026, and it’s a big deal — not just for developers, but for anyone watching the AI industry reshape itself in real time.
Wait, Cursor Makes Its Own Models Now?
Until now, Cursor was essentially a very smart wrapper. You’d open the editor, pick a model (Claude, GPT-4o, Gemini, whatever), and Cursor’s UI and tooling would handle the rest. The models came from outside; Cursor supplied the experience.
Composer 2 changes that equation. It’s Cursor’s own model — built from the ground up on coding-specific data, trained through reinforcement learning on long-horizon tasks, and priced aggressively at $0.50/M input and $2.50/M output tokens.
Aman Sanger, Cursor’s cofounder who leads the research team, was refreshingly blunt about what it can and cannot do:
“It won’t help you do your taxes. It won’t be able to write poems.”
Aman Sanger, Cursor cofounder
That’s not a limitation — it’s a design choice. By focusing entirely on software development, Cursor built a smaller, leaner model that outperforms larger generalist models on coding-specific benchmarks. Think of it as the specialist surgeon versus the do-everything general practitioner debate, but for AI.
The Numbers Are Actually Impressive
Cursor published benchmark results comparing Composer 2 to its predecessors. The jump is substantial:
| Model | CursorBench | Terminal-Bench 2.0 | SWE-bench Multilingual |
|---|---|---|---|
| Composer 2 | 61.3 | 61.7 | 73.7 |
| Composer 1.5 | 44.2 | 47.9 | 65.9 |
| Composer 1 | 38.0 | 40.0 | 56.9 |
SWE-bench Multilingual tests the ability to resolve real-world GitHub issues across multiple programming languages. A score of 73.7 is competitive with frontier models from Anthropic and OpenAI — which is exactly the point.
These improvements come from what Cursor calls their “first continued pretraining run,” which gave a stronger base for reinforcement learning. The model can handle tasks requiring hundreds of sequential actions — the kind of long, complex coding sessions that previously required a lot of human hand-holding between steps.
Why This Matters More Than Just Another Model Drop
Here’s the strategic context that makes this genuinely interesting.
Cursor has over 1 million daily active users. Clients include Stripe, Figma, and 50,000 other businesses. The company has reportedly been in fundraising talks at a valuation around $50 billion. It’s not a scrappy startup anymore — it’s a serious player in the developer tools market.
But there was always a vulnerability: Cursor was dependent on the very companies it was now competing with for users. OpenAI is an investor in Cursor. Anthropic’s Claude models power a significant chunk of Cursor’s usage. Building your own model isn’t just about product quality — it’s about not being at the mercy of your suppliers.
One critical analysis from Vibe Coding (yes, that’s a real website now) put it bluntly: “Cursor was doomed because they don’t control the models, and they can’t aggressively subsidize their own users’ API costs like Anthropic and OpenAI are.”
Composer 2 is, in part, Cursor’s answer to that existential critique.
The Vibe Coding Revolution — A Quick Recap
If you’re not familiar with vibe coding, here’s the short version: it’s a style of programming where you describe what you want in plain language, and the AI writes the actual code. You iterate, you test, you steer — but you’re not necessarily writing every line yourself.
Cursor essentially kicked off this trend when it launched its AI coding assistant in 2023. Within a couple of years, it had a million daily users and was charging companies like Stripe for enterprise access. The vibe coding meme spread from developer Twitter to mainstream tech coverage, and suddenly everyone was building apps without knowing how to code.
Composer 2 is designed to push this further. The model is specifically trained to handle long, multi-step coding tasks autonomously — the kind of work where you say “build me a working authentication system” and then go make coffee while the AI figures out the details.
What About the Price?
Cursor is pricing Composer 2 aggressively. The standard rate is $0.50/M input and $2.50/M output — cheaper than comparable frontier model API access. There’s also a faster variant at $1.50/M input and $7.50/M output, which Cursor claims is still cheaper than other fast models when normalized properly.
For users on individual plans, Composer 2 usage comes out of a standalone usage pool — meaning you get a generous allotment included with your subscription before you start paying per token.
This is classic platform strategy: make the baseline experience cheap enough that switching costs outweigh any savings from going elsewhere. If Composer 2 is good enough and cheap enough, why would you bother plugging in Claude or GPT-4o?
The Bigger Picture: Everyone’s Building Their Own Models Now
Cursor joining the “we make our own AI” club is part of a broader pattern. Every company that built on top of foundation models is now asking the same question: at what scale does it make sense to stop renting and start owning?
The answer for Cursor, apparently, is $50 billion valuation and 1 million daily users.
The interesting ripple effect here is on the model providers themselves. If tools like Cursor (and potentially others — GitHub Copilot, Replit, etc.) start building their own specialized models, the market for “renting” GPT or Claude access for coding tasks shrinks. OpenAI and Anthropic are obviously aware of this, which is why they’ve been releasing their own coding-specific products (Claude Code, anyone?).
It’s becoming a proper arms race, and the battlefield is your IDE.
Should You Try It?
If you’re already a Cursor user, Composer 2 is available today. Cursor has made the fast variant the default option, so you’re likely already using it if you open the app after this update.
If you’re not a Cursor user but you’re curious about AI-assisted coding, this might be a good time to check it out. A tool with a million daily users, clients like Stripe and Figma, and now its own frontier-competitive model isn’t a toy anymore. It’s a development environment that’s taking itself very seriously.
The free tier exists. The model is live. And if the benchmarks hold up in real-world use, this is the first time a coding-specialist AI has credibly claimed to match the generalists at their own game.
The Cat’s Take
Cursor spent three years building on top of other people’s models, and now it’s eating their lunch — or at least trying very hard to. Composer 2 isn’t a moonshot; it’s a disciplined, focused bet that specialization beats generalization in a world where everyone is fighting for the same developer mindshare.
Watch this one. The coding AI space is about to get a lot more interesting.
Sources: Cursor Blog — Introducing Composer 2 | Bloomberg | Economic Times
🐾 Like this kind of AI content? Check out Pudgy Cat’s shop for illustrated books and prints, or browse our books on Amazon KDP if you’re into genre fiction powered by real cat energy.