Composer 2.5
Composer 2.5 adds continued training and reinforcement learning to Kimi K2.5. Training includes targeted text feedback on individual actions and more difficult synthetic coding tasks.
What the release provides
Text feedback targets specific behavior, such as a bad tool choice, rather than relying only on a reward for the complete task. The model is trained to sustain work across many actions and adjust its effort.Composer 2.5 training and base model
How to use it well
Practical guidance. Give the agent an observable outcome and the project’s actual constraints. Inspect the complete change and tool results before accepting it. Long execution does not establish that the requested behavior works.
Speed and price options
Fast is the default Composer 2.5 tier. Cursor describes the same intelligence as the standard tier, with higher speed and different token rates.Current Composer 2.5 context and service tiers
The current product contract supports file edits, search, shell commands and other agent tools. These tools belong to the Cursor environment; their presence does not make Composer a native image generator.Current Composer 2.5 context and service tiers
Practical guidance. Use the fast tier when waiting between interactive changes limits progress. Compare complete task cost, including repeated attempts, before choosing a default for long unattended work.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Provider product |
| Model identifiers | composer-2.5 |
| Context window | 200,000 tokens Current Cursor model context Current Composer 2.5 context and service tiers |
| Standard tier | The standard tier has lower token rates than Fast; both are described as having the same intelligence.Current Composer 2.5 context and service tiers |
Related models
- Composer 1
Cursor (Anysphere) · Composer
First-generation Composer coding-agent model with a 200,000-token context.
Launch date Provider product