Your AI coding assistant just got another contender—from the editor many developers already use to write code. Cursor unveiled Composer 2 today, extending its own coding-model family with a focus on capability and cost. It is competing directly for work that might otherwise go to Claude or 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?
Cursor combines an editor and agent tools with a choice of models, including models from other providers. But Composer 2 is not its first in-house model: the company introduced Composer in October 2025, and had already developed its own completion model for Cursor Tab. The “2” in the name is not merely there for dramatic effect.
The March 19 Composer 2 announcement describes continued pretraining followed by reinforcement learning on extended coding tasks. “Continued pretraining” is not a claim that the entire underlying model was built from scratch. The launch prices start at $0.50 per million input tokens and $2.50 per million output tokens.
In Bloomberg’s launch-day interview, republished by The Economic Times, cofounder Aman Sanger emphasised the coding focus. He joked that it would not do taxes or write poems. The point was its intended job, not a formal test showing that it is incapable of producing a rhyme.
Specialisation is the bet: train a model around the kind of work people actually ask for inside the editor. Whether it beats a particular general model depends on the task and the surrounding tools. Think specialist versus general practitioner, without appointing the benchmark chart as chief medical officer.
The Numbers Are Actually Impressive
Cursor’s published comparison shows a substantial improvement over earlier Composer versions. These are the company’s reported results, not independent tests performed for this article:
| 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 measures software-issue resolution across programming languages. Its score is useful, but it does not make Composer 2 a universal winner over every Anthropic or OpenAI model. CursorBench is Cursor’s own evaluation, and the Terminal-Bench comparisons also involve different agent setups. The footnotes belong with the numbers.
Cursor attributes the gains to continued pretraining and reinforcement learning, and says the model can solve tasks requiring hundreds of actions. That is a capability claim under particular conditions, not a promise that every long coding job will complete without review.
Why This Matters More Than Just Another Model Drop
Here’s the strategic context that makes this genuinely interesting.
Bloomberg’s March 19 report puts Cursor at more than a million daily users and around 50,000 business customers, including Stripe and Figma. It also describes fundraising talks at a roughly $50 billion valuation. A reported negotiating figure is not the same as a completed funding round, but the scale helps explain why model economics matter.
Cursor also occupies an awkward competitive position: its product offers models from firms that sell coding tools of their own. Bloomberg identifies OpenAI as an investor. Developing its own models can give Cursor more control over cost and product choices, though it does not mean all outside dependencies disappear.
A useful way to frame the business question is this: how much control can an editor retain when a supplier also sells the competing editor or agent? That is an interpretation of the competitive setup, not evidence that Cursor was inevitably “doomed”.
Composer 2 continues Cursor’s answer to that problem. It is another step in a model programme that was already under way.
The Vibe Coding Revolution — A Quick Recap
If you’re not familiar with vibe coding, it loosely describes building software by telling an AI what you want and steering the result. The term is often used broadly, though developers who carefully inspect and test AI-assisted code are doing something different from accepting whatever runs.
Cursor’s assistant launched in 2023 and became one of the prominent tools associated with that shift. It did not single-handedly invent natural-language-assisted programming, and the later popularity of the phrase should not be confused with the product’s launch date.
Composer 2 is designed for longer, multi-step coding work. In a good workflow, that means delegating a bounded change and checking the result. “Build authentication while I make coffee” is a tempting demo; deciding whether that authentication is correct still belongs in the afternoon’s plans.
What About the Price?
The standard launch rate is $0.50 per million input tokens and $2.50 per million output tokens. A faster variant costs $1.50 input and $7.50 output. Cursor makes the faster option the default, so the lowest advertised price is not necessarily the rate of the option a user first encounters.
The launch announcement includes Composer usage in an allowance for individual plans. The amount available and what happens after it is used depend on the plan’s terms. “Included usage” does not mean unlimited free access.
The attraction is a combination of cost, speed and integration. If Composer handles a developer’s routine jobs well, switching models may become less necessary. If it struggles with a particular repository, a more expensive option may still be worth it. Your codebase gets a vote.
The Bigger Picture: Everyone’s Building Their Own Models Now
The broader question for companies built around AI is when it makes sense to invest in models as well as the product around them. There is no universal customer count or valuation at which doing so suddenly becomes inevitable.
Cursor is placing that bet while continuing to support outside models. Owning more of the experience and maintaining user choice can coexist.
The potential consequence for model providers is more competition for coding workloads. At the same time, their own coding products compete with tools such as Cursor. It is a two-way contest over both model capability and the environment in which developers use it.
It’s becoming a proper arms race, and the battlefield is your IDE.
Should You Try It?
If you already use Cursor, Composer 2 is available as announced today. Check the selected model and variant before comparing quality or cost; an app update does not necessarily mean every existing conversation has switched models.
If you are curious about AI-assisted coding, a useful trial is a small task in a repository you understand. Look at the proposed changes, run the relevant checks and compare the time spent reviewing them. That tells you more than an impressive company valuation.
Check your plan’s access and allowance before starting a larger job. Composer 2 has encouraging reported results, but this is neither the first specialist coding model nor a reason to treat every generated patch as ready to ship.
The Cat’s Take
Cursor has been combining outside models, its own models and developer tooling for some time. Composer 2 strengthens its attempt to keep more of that work inside its own family. It is a focused bet on a coding workflow—not proof that specialists always beat generalists.
Watch this one. The coding AI space is about to get a lot more interesting.
