Composer 2.5 소개
Composer 2 대비, 특히 장기적인 에이전트형 작업에서 지능과 동작 방식이 크게 개선되었습니다.
Composer
Our most powerful coding model. Built for long-running agents at a fraction of frontier cost.
Announcements
autoinstall로 Composer 부트스트래핑하기
Composer autoinstall이 이전 모델 버전을 활용해 설정과 검증을 자동화함으로써 실행 가능한 RL 환경을 어떻게 부트스트래핑하는지 설명합니다.
Composer 2 기술 보고서
Kimi K2.5 기반의 후속 사전학습부터 실제 Cursor 세션에서의 대규모 RL까지 Composer 2의 학습 과정과 CursorBench, 공개 벤치마크, 인프라를 다룬 arXiv 보고서를 공개했습니다.
실시간 RL로 Composer 개선하기
온라인 강화학습을 Composer에 적용하여 모델 체크포인트를 프로덕션에 서빙하고, 실제 사용자 상호작용을 보상 신호로 사용해 하루에도 여러 번 개선된 체크포인트를 배포합니다.
Use cases
Composer 2.5 pairs frontier-level intelligence with low token cost. It's strong default for every day planning, design, and development work.
With improved effort calibration, Composer 2.5 adjusts how much work it spends based on the complexity of the task. It pushes harder on long-running problems and responds quickly to simpler ones. Popular use cases include:
Fast agentic codingFast agentic
coding
Composer 2.5 keeps pace with interactive sessions. The fast variant returns edits and tool results quickly, so you stay in a tight feedback loop while it handles file edits, terminal commands, and multi-file changes directly in Cursor.
Long-running tasksLong-running
tasks
Trained with reinforcement learning on long-horizon coding tasks, Composer 2.5 sustains effort across extended work. It follows complex instructions reliably, picks the right tools, and stays on the goal through hundreds of tool calls without losing track.
High-volume, cost-sensitive workHigh-volume,
cost-sensitive work
At $0.50/M input and $2.50/M output tokens, Composer 2.5 makes frequent, large-scale use practical. Run it across many sessions, batch routine changes, or power background agents while keeping spend low.
Benchmarks
| Composer 2.5 | Opus 4.7 | GPT-5.5 | Composer 2 | |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 69.3% | 69.4% | 82.7% | 61.7% |
| SWE-Bench Multilingual | 79.8% | 80.5% | 77.8% | 73.7% |
| CursorBench v3.1 (harder tasks) | 63.2% | 64.8% max61.6% xhigh (default) | 64.3% xhigh59.2% medium (default) | 52.2% |
Opus 4.7 and GPT-5.5 use self-reported scores for public evals.
(Above) Composer 2.5 offers frontier-level coding at a fraction of the latency when compared to Opus 4.7 and GPT-5.5 across Terminal-Bench, SWE-Bench, and CursorBench.
Hear from our users
“With Composer 2.5, Cursor gives us the most cost-competitive model at the frontier of intelligence.”
Kyle Cesmat Engineering Manager, Coinbase
“For businesses moving from tokenmaxxing to valuemaxxing, Composer 2.5 could not have come at a better time.”
Austin Ray Staff Software Engineer, Ramp
“Been driving Composer 2.5 for like 3 hours straight and absolutely holy smokes, it is so freaking good!”
Will Newton Principal Product Designer, Amplitude
“Cursor's new Composer 2.5 takes third on the Artificial Analysis Coding Agent Index and is ~10-60x lower cost than the higher-effort Opus 4.7 and GPT-5.5 variants above it.This release puts Composer among the leading coding agent models.”
Artificial Analysis Independent AI Benchmark Provider
“Continuing to test, test, and test some more with Composer 2.5, and I'm getting more and more impressed with each test run.The planning mode really continues to impress me, it is better than Opus 4.7 at planning imo, while being both faster, and using a fraction of the tokens.”
Morgan Linton CTO, Bold Metrics
“Cursor with Composer 2.5 is very cheap and strong: around 8× cheaper than Claude Code and Codex, and scores higher than open-weight models!”
Ibragim Badertdinov Lead ML Engineer, Nebius
Base Model and What's Next
Composer 2.5 is built on the same open-source checkpoint as Composer 2, Moonshot's Kimi K2.5.
Together with SpaceXAI, we're training a significantly larger model from scratch, using 10x more total compute. With Colossus 2's million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.
Kimi K2
Kimi K2.5
Composertraining and RL
0%
(Above) Composer 2.5 is built on an open base with proprietary intelligence: 85% of its compute goes into additional training and RL, rapidly improving its speed, efficiency, and coding abilities.
FAQ
Use Composer 2.5 as your default for hands-on, agentic coding: the tight write, run, and fix loop where it delivers frontier-level quality while using fewer steps, tokens, and cost per task. It's the right call when you want top-tier coding intelligence without the latency, so iteration stays fast.
Composer 2.5 is only available in Cursor. Use it in our desktop app, web, CLI, and mobile. You can also build your own agents on top of Composer 2.5 with our SDK docs.
The standard variant is priced at $0.50/M input and $2.50/M output tokens. Fast mode is priced at $3.00/M input and $15.00/M output tokens. See the model docs for full details.
Composer 2.5 is built on Moonshot's Kimi K2.5, the same open-source checkpoint as Composer 2. Together with SpaceXAI, we're training a significantly larger model from scratch, using 10x more total compute.