Cursor Releases Composer 2.5 to Improve Reliability on Long Running Coding Tasks

CursorCursor

· Updated

Cursor released Composer 2.5, a coding model optimized for sustained performance on complex, multi-step engineering tasks. The update introduces a new reinforcement learning method that provides localized feedback during long trajectories to reduce errors in tool use and communication.

Cursor, an AI-native code editor, released Composer 2.5 to improve reliability on long-horizon agentic tasks. While still based on the Kimi-k2.5 foundation, this version uses targeted reinforcement learning (a training method that aligns model behavior with feedback) to correct specific mistakes—like bad tool calls—within long sequences.
Pricing (input)
$0.50 per million tokens
Pricing (output)
$2.50 per million tokens
Fast tier (input)
$3.00 per million tokens
Fast tier (output)
$15.00 per million tokens
Training compute (future)
10x increase via Colossus 2
Synthetic data scale
25x more tasks than Composer 2

This release addresses the credit assignment problem, where a single reward at the end of a long task makes it hard for models to learn from specific mid-process errors. By scaling synthetic data 25x, Cursor is targeting the multi-day autonomous engineering sessions recently teased for OpenAI's Codex.

You can use Composer 2.5 now within the Cursor editor, with usage limits doubled for the first week. Standard pricing is $0.50 per million input tokens. Looking ahead, a partnership with SpaceXAI aims to train a larger model from scratch using the massive Colossus 2 compute cluster.

View the full update on cursor.com
Cursor
Cursor
@cursor_ai
X

Introducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model. https://t.co/N87ojcXlOC

1.4kretweets13klikes
View on X

Still wondering? A few quick answers below.

Composer 2.5 is an updated AI coding model designed for the Cursor editor that focuses on improving performance during long-running agentic tasks. It is built on the Kimi K2.5 foundation but features architectural refinements that make it more reliable at following complex, multi-step instructions and maintaining intelligence over extended work sessions.

Cursor uses targeted textual feedback to improve model behavior during long trajectories. Instead of providing a single reward at the end of a task, the system inserts specific hints at the exact point where a mistake occurred, such as an invalid tool call. This localized signal helps the model learn precise behavioral corrections without losing the broader objective.

The standard version of Composer 2.5 is priced at $0.50 per million input tokens and $2.50 per million output tokens. There is also a faster variant available that maintains the same level of intelligence for $3.00 per million input tokens and $15.00 per million output tokens, which is the default option in the editor.

The model was trained using 25x more synthetic tasks than the previous version, including techniques like feature deletion where the agent must reimplement functionality in a codebase to pass verifiable tests. This large-scale synthetic data training helps the model find sophisticated solutions and workarounds for difficult real-world engineering problems.

Cursor is collaborating with SpaceXAI to train a significantly larger frontier model from scratch using 10x more total compute than previous efforts. This training will utilize the Colossus 2 cluster, which consists of one million H100-equivalent GPUs, to achieve a major leap in model capability beyond the current Composer 2.5 release.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

See all AI news & updates from Cursor →

Share this update