Cursor introduced Composer 2.5, which the company describes as its "most powerful model yet," claiming it is "more intelligent," better suited to sustained work on long-running tasks, and more reliable at following complex instructions. Cursor also stated that included usage of the model will be doubled for the next week.
In a follow-up post, the company claimed Composer 2.5 is "up to 10x more efficient than similarly capable models." Cursor attributed the update to scaled training, more complex RL environments, and new learning methods, including what it described as text feedback during RL to help assign credit across rollouts spanning "hundreds of thousands of tokens."
Cursor also said Composer 2.5 is built on the same open-source base as Composer 2, Moonshot’s Kimi K2.5. In the same thread, the company added that it is working with SpaceXAI on a much larger model trained from scratch with "10x more total compute," citing Colossus 2’s "million H100-equivalents" and saying the effort could mark "a major leap in model capability."
Replies to the announcement were mixed. Some users asked whether the new model was based on Kimi 2.6 or whether it was open source, while others questioned how it compares on everyday coding tasks and rate limits. One commenter pointed to Cursor’s own benchmark figures, including a claim that Composer 2.5 sits within 1.6% of Opus on CursorBench hard tasks, while another criticized the numbers as self-reported. A separate reply noted a 200k context window as a blocker.
Source: Cursor
