Cursor launches Composer 2.5 with larger RL training and doubled usage

Details

• Cursor introduced Composer 2.5 as its most powerful model yet, aimed at longer-running coding tasks and more reliable instruction following.
• The company says the model is up to 10x more efficient than similarly capable models and will have doubled included usage for the next week.
• Composer 2.5 builds on the same open-source base as Composer 2, Moonshot AI’s Kimi K2.5.
• Cursor says it improved the model by scaling training, creating more complex reinforcement learning environments, and adding new learning methods.
• The company highlights text feedback during RL, which helps assign credit across rollouts spanning hundreds of thousands of tokens.
• Cursor also says it is training a much larger model from scratch with SpaceXAI, using 10x more total compute and Colossus 2’s million H100-equivalents.
• The announcement suggests Cursor is pushing both product iteration and frontier-model training, while keeping the current release tied to its existing coding workflow.

Impact

Composer 2.5 signals that Cursor is competing less as an IDE wrapper and more as a model builder for agentic coding. By emphasizing efficiency, long-horizon task reliability, and temporary usage expansion, it pressures other coding assistants to match both cost and sustained performance. The separate effort with SpaceXAI also points to a broader shift: more software companies are investing in proprietary training stacks rather than relying only on third-party models.