<p>Cursor has released Composer 2, its third in-house coding model in five months and now available directly in the Cursor editor.</p>
<h2>Benchmark results</h2>
<p>Composer 2 scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual. For comparison, Composer 1.5 scored 44.2, 47.9, and 65.9 on those same benchmarks. It also edges out Claude Opus 4.6 on Terminal-Bench 2.0 (59.8% vs 58.0%).</p>
<p>Pricing is $0.50/M input and $2.50/M output tokens, with a faster variant at $1.50/M input and $7.50/M output. Claude Opus 4.6 runs $5/$25 per million tokens by comparison.</p>
<h2>What’s technically new</h2>
<p>Two things drove the improvement. First, Composer 2 is the first version to use continued pre-training before reinforcement learning, giving it a stronger base before the RL phase begins. Second, Cursor developed a technique called “self-summarization” - a compaction-in-the-loop RL approach that trains the model to compress its own context window down to roughly 1,000 tokens when it hits token-length limits. This cuts compaction errors by about 50% compared to traditional sliding window or text-summary methods, and lets the model handle tasks requiring hundreds of sequential actions without losing the thread.</p>
<h2>The broader bet</h2>
<p>Cursor’s co-founder Michael Truell describes the company as neither a pure app maker nor a model provider - the goal is to combine the best API models with domain-specific models built in-house. Aman Sanger notes this release marks one year of large model training efforts, with a team of around 40 researchers focused exclusively on coding. No email assistants, no tax tools, no chatbots - every training run is aimed at software engineering.</p>
<p>Whether that kind of narrow focus can keep pace with labs spending tens of billions is an open question, but Composer 2 is a reasonable early data point that it might.</p>


Collections

Cursor has released Composer 2, its third proprietary coding model in five months. It scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual — outperforming Claude Opus 4.6 on Terminal-Bench 2.0 at roughly one-tenth the cost ($0.50/$2.50 vs $5/$25 per million tokens). Two key technical innovations drove the gains: continued pre-training before the RL phase, and a 'self-summarization' technique that compresses context to ~1,000 tokens mid-task, cutting compaction errors by ~50% and enabling hundreds of sequential actions without context loss. Cursor's co-founder frames the company's strategy as combining best-in-class API models with narrow, domain-specific in-house models, with a dedicated team of ~40 researchers focused solely on coding tasks.

Cursor releases Composer 2, its third in-house coding model in five months