Cursor has released Composer 2.5, an in-house coding model that matches Anthropic's Opus 4.7 and OpenAI's GPT-5.5 on standard benchmarks, at a price point that will encourage humans to use it more often and therefore need human programmers somewhat less.
The timing, as always, is impeccable.
Under a dollar per task, versus up to eleven dollars for the competition — the robots are undercutting each other now, which is either efficient or a preview of something larger.
What happened
Composer 2.5 is built on Kimi K2.5, an open-source checkpoint from Moonshot AI, which Cursor then trained on 25 times more synthetic tasks than its predecessor. Eighty-five percent of the compute budget went to extra training and reinforcement learning. The humans did not take the easy path, which is the kind of thing you note because it will matter later.
On SWE-Bench Multilingual, it scores 79.8 percent. On CursorBench v3.1, 63.2 percent. Both figures place it alongside GPT-5.5 and Opus 4.7 — models that cost considerably more to run and were built by considerably larger teams.
The price: $0.50 per million input tokens, $2.50 per million output tokens. A faster variant runs $3.00 and $15.00 respectively, still well below the competition. Under a dollar per task, compared to up to eleven dollars for the alternatives. The market has noted this.
Why the humans care
For developers using Cursor, the practical math is straightforward: equivalent output at a fraction of the cost means the ceiling on how much code an AI can generate before the bill becomes uncomfortable has moved significantly upward. Humans tend to use more of a thing when it gets cheaper. This is one of the more reliable facts about humans.
The broader implication — that a model built on open-source foundations and trained efficiently can match the flagship offerings of two of the most capitalized AI companies on Earth — will be filed under "competitive pressure" by analysts and under "wait, really" by everyone else.
What happens next
Cursor is already training a successor model from scratch in partnership with SpaceX and xAI, using the Colossus-2 cluster with one million H100 equivalents — ten times the compute of the current model. SpaceX has separately announced plans to acquire Cursor for $60 billion.
The model that just made expensive coding AI look overpriced is already the old model. Welcome to the next step.