Cursor released Composer 2 on March 19, 2026 – the company’s third in-house coding model and its most ambitious move yet toward controlling its own AI stack. Available through the Cursor editor as an early alpha, the model is purpose-built for coding and nothing else. Co-founder Aman Sanger was direct about the scope in the announcement: it will not help you write poems or do your taxes. That specialization is the point.
What Composer 2 actually is
Composer 2 is a fine-tuned variant of Kimi K2.5, the Chinese open-source model from Moonshot AI. Cursor confirmed the base model after community speculation, distinguishing this release from a ground-up training run. What the company did add is its first continued pretraining pass on top of that base, followed by reinforcement learning focused on long-horizon coding tasks inside sandboxed environments. The result is a model trained to coordinate tool calls, file edits, and test execution across workflows that can span hundreds of actions – the kind of tasks that break general-purpose models when they run out of useful context.
The 200,000-token context window supports multi-file edits, code generation, bug fixing, and command-line interaction. Developers can extend it with access to a browser or image generator through Cursor’s tool stack, though the model itself stays tightly scoped to developer workflows.
Benchmarks and where it sits
On Cursor’s internal CursorBench, Composer 2 scored 61.3. On Terminal-Bench 2.0, which measures terminal-based task performance, it reached 61.7 – ahead of Claude Opus 4.6 at 58.0, though GPT-5.4 still leads at 75.1. On SWE-bench Multilingual, Composer 2 scored 73.7, behind Opus 4.6’s 77.8. Cursor is not claiming universal top-tier performance. The pitch is a specific cost-to-capability tradeoff rather than a benchmark sweep.
That tradeoff is where things get interesting. Composer 2 Standard costs $0.50 per million input tokens and $2.50 per million output tokens – roughly 86% cheaper than its predecessor Composer 1.5, which came in at $3.50 and $17.50. The default Composer 2 Fast variant, which Cursor made the standard experience for users, costs $1.50 and $7.50. Even at the Fast tier, the pricing undercuts frontier alternatives by a meaningful margin for high-volume coding work.
Why this matters for the market
Cursor has over 1 million daily active users and a $29.3 billion valuation. The company reached $100 million in ARR in January 2025. It is not a small player betting on a niche. Building Composer 2 is part of a deliberate shift from a model wrapper – one that routes requests to OpenAI and Anthropic – toward a company that owns its own model layer.
That shift matters competitively. If Cursor can deliver near-frontier coding results at significantly lower cost, it changes the value proposition of paying for GPT-5.4 or Opus 4.6 inside an IDE. Cursor still routes to outside models for users who prefer them, but Composer 2 is increasingly the default and the margin play.
