Cursor, an AI-native code editor, announced the release of its new proprietary coding model, Composer 2.5, on May 18, 2026. The company describes Composer 2.5 as a significant advancement over its predecessor, Composer 2, offering enhanced intelligence and a more user-friendly experience for developers.
Composer 2.5 is designed to excel at sustained, long-running coding tasks and is noted for its improved ability to follow complex instructions reliably. According to Cursor, the model aims for “better behavior, not just better benchmarks,” achieved through targeted reinforcement learning (RL) with textual feedback. This approach helps shape specific behaviors such as tool use and communication style, which are crucial for real-world development workflows.
The model’s training leveraged Moonshot’s open-source Kimi K2.5 checkpoint, the same foundation used for Composer 2. However, Composer 2.5 underwent significantly scaled training, incorporating 25 times more synthetic RL tasks than Composer 2. One notable training method involved “feature deletion,” where the model was tasked with reimplementing features in a functional codebase after their removal.
In terms of performance, Composer 2.5 achieved a score of 62 on the Artificial Analysis Coding Agent Index, a 14-point increase from Composer 2’s score of 48. This places it third on the index, positioned behind higher-effort versions of Claude Opus 4.7 and GPT-5.5, which typically incur significantly higher costs. On the SWE-Bench-Pro-Hard-AA benchmark, Composer 2.5’s score of 47% is comparable to Claude Opus 4.7 (max).
Composer 2.5 is also positioned as a cost-effective solution. The standard variant is priced at $0.50 per million input tokens and $2.50 per million output tokens. A faster variant, intended for interactive sessions, costs $3.00 per million input tokens and $15.00 per million output tokens. These prices are notably lower than those of other frontier models, with some reports indicating Composer 2.5 can be 10 to 60 times cheaper per task than competitors like Claude Opus 4.7 and GPT-5.5.
Currently, Composer 2.5 is available as the new default model within the Cursor AI coding assistant. It operates exclusively within the Cursor IDE and CLI, without a public API. Looking ahead, Cursor has announced a partnership with xAI (SpaceXAI) to train a much larger model from scratch, utilizing ten times more total compute, signaling further advancements in their AI coding capabilities.