Mini-Vibe Check: Cursor Bets on Fast and Cheap With Composer 2

Hello, and happy Sunday! Was this newsletter forwarded to you? Sign up to get it in your inbox.

---

Cursor’s new model is competitive but not quite frontier

The benchmarks say Composer 2 competes with the best. Our testers say it’s built for a different kind of developer.

Over the past year, OpenAI and Anthropic have converged on the same vision of AI coding. Claude Code and Codex both optimize for delegation: You describe the problem, the model goes away and solves it, you come back and review.

Cursor, on the other hand, wants the developer in the loop, inside a visual workspace, managing agents in real time. And if you’re running lots of tight cycles instead of delegating a 30-minute task, you need a model that’s fast, cheap, and smart enough.

Composer 2, which Cursor released on Wednesday, aspires to that vision. It comes in two flavors: a standard variant that’s one-tenth the cost of Opus 4.6 and one-fifth the cost of GPT-5.4, and a default “fast” version that’s faster than the fastest Opus model (and still a fraction of the cost).

Cursor says its benchmark scores put Composer 2 at frontier level. In our testing, the benchmarks hold up—but only within Cursor’s ideal workflow.