AI coding startup Cursor released Composer 2 on March 20, a specialized programming model with a 200,000-token context window built exclusively for multi-file edits, refactoring, and long-running agent tasks inside its editor. The model scored 61.3 on CursorBench and 73.7 on SWE-bench Multilingual, outperforming Claude Opus 4.6 across many programming tasks while trailing only GPT-5.4's top configurations.

Composer 2 is a fine-tuned variant of Chinese open-source model Kimi K2.5, trained solely on coding data. Standard pricing starts at $0.50/$2.50 per million input/output tokens — roughly 86% cheaper than its predecessor Composer 1.5. A faster premium tier runs at $1.50/$7.50 per million tokens.

The launch signals a shift in the AI coding market: purpose-built models trained on narrow datasets can now compete with general-purpose giants at dramatically lower cost. Cursor, valued at $10 billion after its last funding round, is betting that developers will prefer a cheaper, coding-optimized model over more expensive all-purpose alternatives from OpenAI and Anthropic.