Cursor's Composer 2 Beats Claude Opus 4.6 on Coding Benchmarks — at 86% Lower Cost
Cursor launched Composer 2 on March 19, a fine-tuned coding model built on Kimi K2.5 that outperforms Anthropic's flagship on CursorBench and SWE-bench Multilingual while cutting input token costs from $3.50 to $0.50 per million.
Cursor shipped Composer 2 on March 19, and the numbers are hard to dismiss. The model scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual — all ahead of Claude Opus 4.6, which the company lists as a direct comparison in its published benchmarks. GPT-5.4 still leads Terminal-Bench at 75.1, but Composer 2 beats it on SWE-bench Multilingual.
The cost differential is where this gets interesting. Composer 2 Standard runs at $0.50/M input and $2.50/M output. Composer 1.5, released six weeks ago, was $3.50/$17.50. That’s an 86% reduction in input cost and a similar drop on output. Composer 2 Fast, for latency-sensitive tasks, runs $1.50/$7.50 — still far below what frontier labs charge for comparable output quality on code.
Cursor’s bet here is architectural, not just benchmark-chasing. Composer 2 is a fine-tuned variant of Kimi K2.5 — a Chinese open-source base model — continued-pretrained and reinforcement-learned on long-horizon coding tasks inside the Cursor environment. The 200,000-token context window is tuned for tool use, multi-file edits, and terminal operations. Cursor didn’t try to build a general-purpose model. It built a model for one job and trained it relentlessly for that job.
The business logic tracks. Anysphere, the company behind Cursor, doesn’t need Composer 2 to write poetry or summarize PDFs. It needs the model to close diffs, navigate unfamiliar codebases, and execute hundreds of sequential tool calls without losing context. By narrowing scope, they shrank the model, cut inference cost, and apparently improved benchmark performance against much larger general-purpose systems.
This matters beyond Cursor’s own product. It’s a proof of concept for task-specific fine-tuning over general frontier reasoning: you don’t always need the smartest model in the room, you need the most focused one. If Composer 2 performance holds up in production — which early developer reports suggest it does — expect other coding tools to follow the same playbook.
The practical implication for developers: running Cursor’s agentic tasks at scale just got significantly cheaper, with better context retention on complex multi-file projects. For anyone doing serious automated coding work, those two things in combination are more valuable than raw benchmark points.