Composer 2.5 Fast

What is Composer 2.5 Fast?#

Composer 2.5 Fast is the low-latency default variant of Composer 2.5, Cursor's coding model released on May 18, 2026. It runs the same model as standard Composer 2.5 on faster serving, so tokens start streaming sooner and interactive sessions feel closer to real time. The family is built on Moonshot's Kimi K2.5 open checkpoint, then post-trained with reinforcement learning on real software tasks. On Unifically it runs as cursor/composer-2.5-fast through the OpenAI-compatible chat completions endpoint. Use it when a person sits in the loop: edits, debugging, and short agent runs where a slow response breaks the flow.

Key features of Composer 2.5 Fast#

Same intelligence, faster serving

Fast runs the same Composer 2.5 model with quicker response starts. Cursor makes it the default variant for interactive work and keeps the standard variant for background jobs.

Near-frontier coding scores

Composer 2.5 scores 79.8% on SWE-bench Multilingual and 69.3% on Terminal-Bench 2.0, within a point of Claude Opus 4.7 on both.

Low cost per finished task

On CursorBench v3.1 hard tasks, Composer 2.5 scores 63.2%, above the default settings of Opus 4.7 and GPT-5.5, at a lower average cost per task.

Trained on real software work

The 2.5 release used 25x more synthetic tasks than Composer 2 plus targeted feedback during training, which shows up as steadier long-task follow-through and better instruction following.

Agent tool use

Multi-file edits, terminal commands, and codebase-wide search are the tasks the model was tuned for, so it holds up in agent loops, not just single completions.

Best for#

Use cases#

Build an in-editor assistant that handles small requests on demand: explain an error, patch a function, write a test, each with a short wait. Use it as the interactive model in a coding agent, with tool calls for file edits and terminal commands, and hand long unattended runs to the standard variant. It also fits chat products for developers, where response speed decides whether people keep using the tool. For CI helpers, it can triage failing tests and draft small fixes quickly enough to sit inside the feedback loop instead of after it.

Limitations#

Terminal-heavy work trails the best frontier result on that axis. Composer 2.5 scores 69.3% on Terminal-Bench 2.0 against 82.7% for GPT-5.5.

CursorBench v3.1 numbers come from Cursor's own harness, so treat them as directional rather than independently verified.

There are no reasoning-effort or thinking controls. The model runs one way, so tasks that need deep multi-step reasoning outside of code may do better on a thinking-capable model.

Fast costs more per token than standard Composer 2.5. Long background jobs are cheaper on the standard variant.

Composer 2.5 Fast vs Composer 2.5#

Both variants run the same model; the difference is serving. Fast streams tokens sooner and is the default for interactive work, while the standard variant costs less per token and fits batch jobs. An independent 110-run skills eval timed Fast at 59 seconds per task on average against 87 for the standard variant, with quality scores within a point of each other. If someone is watching the screen, use Fast. If the job runs unattended, the standard variant is the cheaper choice.

When to use Composer 2.5 Fast#

Use Composer 2.5 Fast when a person is in the loop and latency shapes the experience: in-editor assistants, debugging chats, and short agent runs. Pick standard Composer 2.5 for unattended batch work where per-token cost matters more than wait time. Move to a heavier general model when the task leans on deep reasoning outside of code.

API examples#

Call Composer 2.5 Fast from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/cursor/composer-2.5-fast.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "cursor/composer-2.5-fast",
    "messages": [
      { "role": "user", "content": "Explain this error and propose a fix: TypeError: Cannot read properties of undefined." }
    ]
  }'

The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.