Cursor dropped Composer 2 the day before yesterday, claiming it’s stronger than Claude Opus 4.6 and 10 times cheaper. Then, while a developer was debugging the API, they spotted this line in the request: accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast Shortly after, 月之暗面 (Dark Side of the Moon) proactively tweeted to confirm: Composer 2 is trained using Kimi-k2.5 as the base model, with Cursor adding further pretraining and extensive RL training on top. That tweet got 17k likes. Cursor’s VP of Developer Experience, Lee Robinson, came out to clarify: yes, the base model is open source, but in the final product, only about 1/4 of the compute power comes from the base model—the rest is Cursor’s own work. Not exactly lying, just a bit of strategic wording. What’s interesting is this supply chain: Cursor, valued at $30 billion and going head-to-head with Anthropic and OpenAI, is relying on a core agent model built by the Chinese company Kimi. Composer 2 scored 61.7% on terminal-bench, compared to Opus 4.6’s 58%, and it’s only 1/10th the price. When you’re running an agent in Cursor, it’s basically Kimi helping you write code. Gotta say, Chinese models are bringing the heat.