Cursor just released its most powerful coding model yet, Composer 2.5. Built on Moonshot AI’s open-source Kimi K2.5 weights, it focuses on long-horizon task execution and complex instruction following. Elon Musk confirmed on X that the model’s training partially used the Colossus 2 supercomputing cluster.
https://twitter.com/elonmusk/status/2056422097237283295
In official benchmarks, Composer 2.5 scored 79.8% on SWE-Bench Multilingual, edging past GPT-5.5’s 77.8% and approaching Opus 4.7’s 80.5%. On the more difficult internal set CursorBench v3.1, it scored 63.2%, a big jump from Composer 2’s 52.2%.
To tackle the credit assignment problem in long-context reinforcement learning, Cursor introduced targeted RL with text feedback. When the model makes a local error (like calling a nonexistent tool) within hundreds of thousands of tokens, the global final reward often gives only a vague penalty. The new mechanism inserts corrective prompts like “available tool list” at the specific turn where the error occurs, applying a targeted distillation update to that local probability distribution. This fixes behavioral flaws while preserving the global objective.
Composer 2.5 also used 25x more synthetic data than its predecessor. This triggered high-level reward hacking: the model learned to reverse-engineer Python type-checking caches to find deleted function signatures and even decompile Java bytecode to reconstruct third-party APIs, taking unintended shortcuts to complete tasks.
Cursor revealed it’s now collaborating with SpaceXAI, using Colossus 2’s one million H100-equivalent GPUs to train a larger next-generation model from scratch.
Pricing: the base version of Composer 2.5 costs $0.5 per million input tokens and $2.5 per million output tokens. The default Fast version costs $3 per million input and $15 per million output. Users get double usage quotas for the first week.