Cursor Composer 2 is now available, and it marks a significant step forward in AI-powered coding tools. Built by the team at Cursor, this new model delivers frontier-level performance on coding benchmarks while keeping costs low enough to make it genuinely practical for everyday development work.
What Is Cursor Composer 2?
Composer 2 is Cursor’s latest proprietary AI model, designed specifically for coding tasks. It replaces Composer 1.5 and brings major gains across every benchmark the team tracks. The model can handle long-horizon tasks requiring hundreds of sequential actions, something earlier versions struggled with. Think of it as the difference between a model that writes a function and one that builds an entire feature from scratch.
Benchmark Results: How Does It Stack Up?
The numbers tell a clear story. Across three major benchmarks, Composer 2 outperforms its predecessors by a wide margin. Here is how the models compare:
| 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 |
That jump from 44.2 to 61.3 on CursorBench is not incremental. It reflects a fundamentally different training approach. Cursor ran its first continued pretraining run to build a stronger base model, then applied reinforcement learning on long-horizon coding tasks on top of that foundation.
Cursor Composer 2 Pricing: Competitive and Transparent
Cost is where Cursor Composer 2 makes a strong case for itself. The standard version is priced at $0.50 per million input tokens and $2.50 per million output tokens. For developers who need speed, a fast variant is available at $1.50/M input and $7.50/M output. Cursor says this fast variant is still cheaper than comparable fast models from other providers. On individual plans, Composer usage draws from a dedicated usage pool with generous included limits, so most developers will not hit a wall quickly.
How It Was Trained
Cursor published a technical report alongside the launch, detailing the training methodology. The key innovation is a two-stage process: first, a continued pretraining run that gives the model a much stronger coding foundation than a general-purpose base model would have. Second, reinforcement learning applied specifically to long-horizon coding tasks. This combination is what allows Composer 2 to plan and execute across hundreds of steps without losing context or making compounding errors. For developers interested in the technical details, the full report is available on the Cursor blog.
Who Should Use It?
Any developer already using Cursor will want to switch to Composer 2 immediately. The performance gains are substantial, and the pricing is not punishing. For teams building complex software applications, the ability to handle multi-step tasks autonomously could meaningfully reduce the time spent on repetitive scaffolding and boilerplate work. The model is also available in the early alpha of Cursor’s new interface, giving users a chance to test it in a fresh environment before it becomes the default experience.
Cursor is positioning Composer 2 as the fast default option going forward. Given the benchmark scores and the pricing structure, that is a reasonable call. The AI coding tools space is moving fast, and this release puts Cursor back at the front of the pack.
