Cursor has released Composer 2.5, the native model for its agentic software-development environment, designed and tuned specifically for developer workflows. Composer 2.5 is an evolution of Composer 2 (released in March) and, like its predecessor, is built on Moonshot’s open-weights Kimi K2.5.
Technical specs and capabilities
- Input/output: text in (up to 200,000 tokens), text out; image output available via tool calls.
- Architecture: mixture-of-experts transformer — 1.04 trillion parameters, with 32 billion active parameters per token.
- Features: function calling, reasoning, context caching.
Composer 2.5 is intended to be used together with tooling for code execution and editing, and was fine-tuned with a harness aligned to Cursor CLI.
Training and fine-tuning approach
The model follows the training recipe Cursor published for Composer 2. Starting from Kimi 2.5’s open weights, the team performed additional pretraining on a large code dataset and then applied reinforcement learning in a simulated agentic environment using tools that mirror Cursor CLI.
During RL, the model was rewarded for success as well as concision and elegance of outputs. The team introduced two noteworthy changes:
- Alongside reward signals, they injected text feedback into the model’s context window. If the model made an incorrect tool call, the context would include suggestions for better tool calls and the correct output to teach the model.
- They fine-tuned the model on 25 times as many synthetic tasks as used for Composer 2, focusing on harder, multi-step problems (for example, a task that asks to remove a feature from an application paired with code changes, cleanup, and tests to confirm the app still works). The team did not disclose which model generated those synthetic tasks.
Performance and benchmarks
Composer 2.5 competes with leading generalist models on multiple coding benchmarks but typically ranks just behind the top performers:
- On the Artificial Analysis Coding Agent Index (a composite of SWE-Bench-Pro-Hard-AA, Terminal-Bench v2, and SWE-Atlas-QnA), Cursor CLI running Composer 2.5 scored 63, behind Claude Code with Claude Opus 4.7 Max (67) and Codex with GPT-5.5 at xhigh reasoning (65). It outperformed those models when they were configured at lower reasoning levels and those same models running inside Cursor CLI.
- In Artificial Analysis’s mean time per task measurement, Cursor CLI with Composer 2.5 averaged 6.7 minutes per task. Claude Code with Claude Opus 4.8 at medium reasoning took 8.8 minutes, while Max reasoning took over twice as long (17.7 minutes).
- On mean cost per task, Cursor CLI Composer 2.5 in fast mode averaged $0.44, versus $4.14 for Claude Code with Claude Opus 4.7 Max.
- On Cursor’s own CursorBench, which is designed to better reflect terse agentic inputs and harder problems, Composer 2.5 scored 63.2%, slightly behind Claude Opus 4.7 (64.8%) and GPT-5.5 (64.3%) at their highest reasoning settings, but ahead of those models at their default settings (61.6% and 59.2%).
Overall, Composer 2.5 typically runs faster and cheaper than comparable models, generally exceeding only DeepSeek V4 Pro in cost among tested systems.
Availability and pricing
Composer 2.5 is available through Cursor’s IDE and via Cursor’s service. Token pricing is as follows:
- Base: $0.50 / $0.20 / $2.50 per million input / cached / output tokens;
- Fast mode: $3.00 / $0.50 / $15 per million input / cached / output tokens.
Subscription plans for individuals, teams, and enterprises start at $20 per month.
Background: SpaceX and Cursor’s roadmap
In April, SpaceX obtained the right to acquire Cursor for $60 billion or to pay $10 billion for their joint work, as part of a wider partnership. As part of that collaboration Cursor will train models on SpaceX hardware and is training new models from scratch, which could reduce its reliance on Moonshot’s open-weights offerings over time.
Why this matters
Cursor’s Composer 2.5 underscores the growing importance of harness engineering—the integration of model and tooling that enables agentic tasks—alongside the capabilities of base models themselves. By developing a model specifically tuned within its coding harness, Cursor aims to combine model strengths with tightly integrated software to deliver faster and lower-cost solutions for developers.
Specialist, task-focused models remain relevant in the agentic era, particularly when they are trained and tuned to work closely with the surrounding tools and workflows that developers actually use.
