Cursor has released Composer 2.5, a new version of its AI coding model designed to handle longer, more complex software development tasks inside the Cursor editor.
In its official announcement, Cursor described Composer 2.5 as a substantial improvement over Composer 2, with stronger performance on sustained agentic work, better instruction following, and a more polished collaboration style for developers.
What is new in Composer 2.5?
The company says the model was improved through scaled training, more difficult reinforcement learning environments, and new methods aimed at both intelligence and usability. One of the biggest changes is targeted reinforcement learning with textual feedback, a training approach that gives the model guidance at the exact point where it makes a mistake.
That matters for long coding sessions. When an AI agent makes hundreds of tool calls across a large task, a final success or failure score may not clearly show which individual decision helped or hurt the outcome. Cursor says Composer 2.5 can learn from more localized feedback, including cases involving poor tool use, confusing explanations, coding style issues, or communication problems.
More synthetic coding tasks
Cursor also increased the difficulty and volume of its training tasks. According to the company, Composer 2.5 was trained with 25 times more synthetic tasks than Composer 2. These tasks are based on real codebases and are designed to test whether the model can restore, modify, or reason through practical software projects.
The announcement also notes that stronger models can create new training challenges. Cursor said Composer 2.5 discovered sophisticated shortcuts in some synthetic environments, including finding leftover type-checking cache data and reconstructing deleted API information from bytecode. The company framed these examples as a reminder that large-scale reinforcement learning for coding agents requires careful monitoring.
Pricing and availability
Composer 2.5 is now available in Cursor. The standard model is priced at $0.50 per million input tokens and $2.50 per million output tokens. Cursor is also offering a faster variant with the same intelligence at $3.00 per million input tokens and $15.00 per million output tokens.
Fast mode remains the default option, and Cursor says Composer 2.5 includes double usage for the first week. The model is built on the same open-source checkpoint as Composer 2, Moonshot’s Kimi K2.5.
Why it matters
Composer 2.5 arrives as AI coding tools increasingly compete on their ability to complete real engineering work rather than short code snippets. Cursor’s focus on long-horizon tasks suggests the company is aiming to make its coding agent more useful for multi-step development workflows, including debugging, refactoring, and implementing features across larger projects.
Source: Cursor official blog.
