Cursor Ships Composer 2 With Major Benchmark Gains

Cursor released Composer 2, their in-house coding model, with benchmark jumps across the board - 61.3 on CursorBench (up from 38.0 for v1), 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual. The gains come from continued pretraining on long-horizon coding tasks that require hundreds of sequential actions. Pricing sits at $0.50/$2.50 per million tokens for standard, with a faster variant at $1.50/$7.50 that ships as default. Cursor training their own models rather than just wrapping Claude or GPT is the real story - it's a bet that coding-specific pretraining and RL can outperform general-purpose frontier models on the tasks that matter inside an editor. xAI just hired two Cursor engineers weeks ago, which makes the timing of this release feel deliberate.