Robert Youssef on X: "Cursor just built their own AI model. It beats Claude Opus 4.6 on real-world coding tasks and costs a fraction of the price. > Composer 2 scores 61.3% on CursorBench versus Opus 4.6's 58.2%. Same tasks. Cheaper inference. Built specifically for the work developers actually do.

Cursor just built their own AI model. It beats Claude Opus 4.6 on real-world coding tasks and costs a fraction of the price. > Composer 2 scores 61.3% on CursorBench versus Opus 4.6's 58.2%. Same tasks. Cheaper inference. Built specifically for the work developers actually do. > Cursor didn't benchmark Composer 2 on SWE-bench. They built their own evaluation suite from actual coding sessions run by their engineering team. > CursorBench tasks have a median of 181 lines changed per task versus 7 to 10 lines for SWE-bench Verified. Prompts are deliberately underspecified, averaging 390 characters versus 1,185 to 3,055 characters for public benchmarks. The agent receives a terse bug report, a reference to production logs, and a codebase. No hand-holding. No narrow specification. Figure out what's broken and fix it. That's what real software engineering looks like. That's what Composer 2 was trained on. > The architecture is a 1.04 trillion parameter Mixture-of-Experts model with 32 billion active parameters, starting from Kimi K2.5 as the base. Cursor ran continued pretraining on a large code-dominated data mix, extending context to 256k tokens, then applied large-scale reinforcement learning on tasks that directly emulate real Cursor sessions the same tools, the same harness, the same environment the deployed model runs in. The RL training distribution covers debugging, new features, refactoring, codebase understanding, testing, code review, optimization, DevOps, and migration. > Not just bug fixes. The full spectrum of what developers actually ask an agent to do. > The infrastructure finding is what separates this from standard fine-tuning. Cursor trained with weight updates mid-rollout inference workers update weights while a trajectory is still being generated, so later tokens in a rollout are less off-policy. They replay MoE expert routing from inference to training to eliminate numerical disagreement between the two forward passes. They compress weight updates to diffs rather than full parameter uploads, getting the 1T parameter model's updates down to a few gigabytes per sync. The entire RL pipeline ran across 3 GPU regions and 4 CPU regions simultaneously. The full benchmark results: → Composer 2 on CursorBench: 61.3% vs Opus 4.6's 58.2% → Composer 2 on SWE-bench Multilingual: 73.7% vs Opus 4.6's 75.8% → Composer 2 on Terminal-Bench: 61.7% vs Opus 4.6's 58.0% → 37% relative improvement over Composer 1.5 on CursorBench → 61% improvement over Composer 1 → CursorBench tasks: median 181 lines changed vs 7-10 for SWE-bench → Inference cost: Pareto-optimal accuracy competitive with frontier, cost comparable to smaller models → Both average performance AND best-of-K improve during RL training no diversity tradeoff > The result challenges a core assumption about frontier AI. The default strategy has been to use the biggest general-purpose model available and prompt it well. Cursor's finding is that a model trained specifically on your domain, your tools, your task distribution, and your harness outperforms general models that cost significantly more to run. The gap between Composer 2 and Opus 4.6 on CursorBench is not marginal. It's consistent across the full evaluation set. And Composer 2 gets there at lower inference cost. > OpenAI suspended SWE-bench Verified reporting after finding evidence that frontier models could generate gold patches from memory. Haiku 4.5 scores 73.3% on SWE-bench Verified nearly identical to GPT-5's 74.9% despite the two models performing very differently on broader task distributions. The benchmarks that the industry uses to compare coding models are contaminated, narrow, and increasingly disconnected from what developers actually need. Cursor built a benchmark that isn't. Then they built a model that wins on it. Domain specialization just beat general intelligence. And it did it cheaper.