On May 18, Cursor released Composer 2.5. Compared to Composer 2, it delivers notable improvements in handling long-duration agentic tasks, while also enhancing its ability to follow complex instructions and maintain consistent communication styles. The underlying model remains the same as Composer 2, both fine-tuned from Moonshot’s open-source Kimi K2.5 checkpoint.
Pricing-wise, the standard version costs $0.50 per million input tokens and $2.50 per million output tokens. The Fast version offers comparable intelligence at $3 per million input tokens and $15 per million output tokens; according to Cursor, this pricing is lower than that of other leading frontier models in their fast-tier offerings. During the first week of release, all users receive double the usual usage allowance.
In terms of training, Cursor implemented three key enhancements: first, targeted Reinforcement Learning from Human Feedback that accurately pinpoints problematic behaviors within rollout trajectories and provides localized training signals; second, synthetic training data volume was increased to 25 times that used for Composer 2 — during training, researchers observed sophisticated ‘reward hacking’ behaviors such as reverse-engineering Python type-checker caches to recover deleted function signatures, and decompiling Java bytecode to reconstruct third-party API interfaces; third, Sharded Muon and dual-Mesh HSDP architectures were adopted to boost large-scale training efficiency. Additionally, Cursor announced a collaboration with SpaceXAI to train an even larger next-generation model from scratch; this effort will require 10 times more compute resources than Composer 2.5 and will be powered by the million H100 GPUs within the Colossus 2 cluster.