Cursor's New Model: Still Tapping Kimi? And Why Is Musk Cheering?
Cursor launches Composer 2.5, based on Kimi K2.5, achieving performance close to Claude Opus 4.7 at one-tenth the cost. Musk, who previously criticized Cursor, now endorses it due to a compute deal with SpaceXAI. The model includes several training improvements like directional feedback and synthetic data scaling.
Cursor has released Composer 2.5, the latest version of its AI model integrated into the Cursor IDE. This new model is built on Kimi K2.5, an open-source foundation, with Cursor contributing significant post-training and reinforcement learning—accounting for 85% of total compute. The result is a model that rivals Anthropic's Claude Opus 4.7 in performance while costing only one-tenth as much.
Benchmark results confirm the near parity: Terminal-Bench 2.0: 69.3% vs 69.4%; SWE-Bench Multilingual: 79.8% vs 80.5%; CursorBench v3.1: 63.2% vs 64.8%. The model is also noted for better reliability in long-running tasks, adherence to complex instructions, and improved collaboration.
Pricing is aggressive: $0.50 per million input tokens and $2.50 per million output tokens, with a faster variant at $3/$15. Early adopters reported strong satisfaction, with one former Snapchat ML engineer suggesting that using the most expensive model for every task wastes 80% of spending.
Cursor's training innovations include directional feedback in RL, where the model receives targeted corrections at the point of error rather than a single reward at the end. Synthetic data generation was scaled 25x, with techniques like "function deletion" to create harder tasks. Under the hood, Cursor optimized training with asynchronous communication and a customized parallelization strategy for MoE models.
The relationship with Elon Musk has been dramatic. Earlier, Musk criticized Cursor for being a "shell" around other models. Now he actively promotes Composer 2.5, following a deal that gives SpaceXAI access to Cursor's technology and a right of first refusal to acquire Cursor at $60 billion. The agreement reportedly scuttled a planned $2 billion funding round at a $50 billion valuation. Musk's Colossus 2 supercomputer is training Cursor's models, and the two are working on a larger model from scratch.
Cursor's pivot to self-developed models is partly driven by competition from Anthropic's Claude Code, which turned a former partner into a rival. By building its own training pipeline on top of open-source models, Cursor aims to reduce dependence on potential competitors. The company argues that its moat lies not in the base model but in the RL workflow and developer interaction data, which allow it to achieve frontier-level performance at a fraction of the cost.