Cursor Releases Composer 2.5 with Enhanced Training and Behavioral Improvements

Cursor has released Composer 2.5, announced in a company blog post. The new version offers major improvements in intelligence, behavior, and reliability when working on complex or long-running tasks.

Composer 2.5 introduces targeted reinforcement learning with textual feedback to help the model correct specific behaviors during training. This approach provides localized hints to refine model responses and address issues such as incorrect tool calls or unclear communication. The update also expands synthetic data generation, with the model trained on 25 times more synthetic coding tasks than its predecessor.

The model’s training stack includes modifications to the Muon optimizer and dual mesh HSDP for more efficient distributed computation. These methods help maintain high performance while managing large-scale model parameters across multiple GPUs. Composer 2.5 is built on the same open-source checkpoint as Composer 2, Moonshot’s Kimi K2.5.

Composer 2.5 is available at $0.50 per million input tokens and $2.50 per million output tokens, with a faster variant priced at $3.00 and $15.00 respectively. The faster version is the default option and includes double usage for the first week.