Cursor Launches Composer 2 — Its Own Frontier Coding Model That Beats Claude Opus 4.6

Cursor stopped being an IDE wrapper and became a model company.

Cursor shipped Composer 2, a proprietary coding model trained via reinforcement learning on long-horizon coding tasks. On CursorBench — their own benchmark, caveats acknowledged — it scores 61.3, beating Claude Opus 4.6's 58.2. It supports 200K context, runs via CLI, and costs $0.50/M input and $2.50/M output tokens. That pricing is an order of magnitude below frontier API rates.

This is the first time a major IDE vendor has shipped its own frontier-competitive coding model rather than routing to Anthropic, OpenAI, or Google. The strategic shift is significant: Cursor previously differentiated on UX, context management, and IDE integration. Now it differentiates on the model itself. Every other coding tool that relies on third-party APIs just lost a structural advantage — Cursor controls both the interface and the intelligence.

The pricing deserves attention. At $0.50/$2.50, Composer 2 undercuts Claude Sonnet 4.6 by roughly 6x on input and 3x on output. For high-volume agentic coding workflows where API costs compound — multi-file refactors, long debugging sessions, CI integration loops — the cost difference is material. Teams running thousands of agent interactions per day would see monthly bills drop significantly.

The benchmark question matters. CursorBench is Cursor's own evaluation suite, and self-reported benchmarks should always carry asterisks. But the directional claim — that a model trained specifically for coding tasks via RL on coding trajectories can beat a general-purpose frontier model on coding — is architecturally plausible. Specialized training on the target distribution should win against general capability, all else being equal.

The competitive read: Anthropic and OpenAI now face a customer that's also a model competitor. Cursor's 2M+ developers represent both a user base and a training data flywheel. Every coding session generates trajectories for RL training. The more people use Cursor, the better Composer gets. That's a loop Anthropic can't replicate from API logs alone.