Cursor drops Composer 2 coding model with frontier performance

If you've fired up Cursor recently, you probably noticed the pitch for Composer 2, its new coding model promising frontier-level intelligence at a fraction of the cost.

Then came the plot twist: a developer named Fynn spotted code that seemingly left Moonshot AI's open-source model ID unscrubbed, joking that Cursor had shipped Kimi 2.5 without even renaming it.

Cursor soon stepped forward to confirm the lineage, sparking a fresh debate over what makes an AI model proprietary.

A Chinese base with high-compute fine-tuning

Cursor claims significant performance leaps for Composer 2 over Composer 1.5. The model scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual.

It is priced aggressively at $0.50 per million input tokens and $2.50 per million output tokens, with a faster variant at $1.50 input and $7.50 output that Cursor makes the default. Individual paid plans include Composer in their first-party usage pool.

When caught, Cursor developer education VP Lee Robinson admitted Composer 2 started from an open-source base, but noted that roughly three-quarters of the compute went into Cursor's own continued pretraining and reinforcement learning for multi-step agent actions.

Moonshot AI chimed in to congratulate Cursor, confirming an authorized commercial partnership through Fireworks AI. Cursor co-founder Aman Sanger also acknowledged that omitting the Kimi attribution in the launch blog was an oversight.

Why it matters

For developers, Composer 2 is a no-brainer to try today. Getting frontier coding agent performance at fifty cents per million input tokens drops the friction of letting an AI grind through hundreds of terminal commands and file edits.

For the wider industry, this cuts through the hype around proprietary full-stack supremacy. Cursor is a $29.3 billion juggernaut reportedly doing over $2 billion in annualized revenue, yet it chose an open base from a Chinese lab rather than spending hundreds of millions building a foundation model from scratch.

We think Cursor's technical strategy is brilliant: foundation models are becoming a commodity, while proprietary interface data and reinforcement learning loops are where agentic value actually lives. But hiding the base until an independent developer called them out was an unforced error. In an open-source world, transparency builds far more trust than pretend-native pedigree.