Cursor just crushed Claude Code

Summary

Key Points

• —Codium (Cursor) released Composer 2.5, a small yet highly effective AI model specifically designed for coding tasks, which the presenter believes is being overlooked. 0:06
• —Model pricing involves input and output token costs, but Composer 2.5 stands out for its extreme token efficiency, generating significantly fewer tokens to solve problems compared to larger, more expensive models. 6:06
• —Cursor faces a competitive disadvantage from major labs like Anthropic and OpenAI, which heavily subsidize their models for users, making Cursor's API usage more expensive for its own platform. 8:35
• —Cursor leverages its extensive user interaction data, particularly chat histories and feedback on agent-generated code, to train smarter models and offset the high costs of using external APIs. 9:03
• —Composer 2.5 is currently only usable within the Cursor platform, making external benchmarking difficult, though Cursor's internal "Cursor Bench" shows impressive results. 11:37
• —On Cursor's internal benchmark, Composer 2.5 scores comparably to high-end models like GPT 5.5 and Opus 4.7 but at a fraction of the cost, making it exceptionally cost-effective for coding tasks. 12:20
• —Cursor achieved Composer 2.5's intelligence through innovative training techniques, including targeted Reinforcement Learning (RL) with textual feedback to correct specific behaviors and the generation of 25x more complex synthetic tasks. 13:11
• —Composer 2.5 is particularly powerful for developers engaged in collaborative, back-and-forth workflows with the AI, where its speed and intelligence enhance productivity in real-time coding scenarios. 32:09
• —The significant compute power required for Composer 2.5's advanced training was enabled by Cursor's collaboration with SpaceX AI, hinting at future large-scale model development. 36:34
• —Cursor is training a significantly larger model with SpaceX AI, using 100x more compute than Kimmy, which could potentially enable them to leapfrog current state-of-the-art AI models for code in the near future. 36:34