Cursor has shipped Composer 2, the company’s first internally built AI model designed for multi-step coding tasks. Until now, Cursor relied entirely on third-party models like Claude and GPT for its code generation features. Composer 2 changes that by offering a purpose-built model optimized for the way developers actually use Cursor: long editing sessions that span multiple files and require sustained context.
The shift matters because general-purpose models waste tokens on capabilities that coding workflows never need. Composer 2 strips those out and focuses compute on code understanding, file navigation, and incremental edits across large projects.
Why Cursor Built Composer 2 for Agentic Coding
- Purpose-built for multi-file, multi-step coding sessions that stretch across hours
- Reduced inference costs compared to routing every request through a third-party API
- Tighter integration with Cursor’s editor context, including open tabs, terminal output, and git history
- Optimized for agentic workflows where the model plans, edits, tests, and iterates without constant prompting
How Composer 2 Handles Long-Horizon Tasks
Most coding AI tools work well for short tasks: write a function, fix a bug, generate a test. They struggle when the task involves coordinating changes across five or ten files while maintaining consistency. Composer 2 was trained specifically on these long-horizon scenarios.
In Cursor’s internal benchmarks, Composer 2 completed multi-step refactoring tasks with 23% fewer failed attempts than Claude Sonnet 4.6 on the same prompts. The model also used 40% fewer tokens per completed task, which translates directly to lower API costs for teams on usage-based plans.
Composer 2 completed multi-step refactoring tasks with 23% fewer failed attempts and 40% fewer tokens than comparable third-party models in Cursor’s benchmarks.
What Developers Should Expect from Composer 2
Composer 2 is available now in the latest Cursor update. You can select it as your default model or use it alongside Claude and GPT models depending on the task. For quick single-file edits, the third-party models still perform well. For anything that requires planning across multiple files, Composer 2 is worth testing as your primary option.
Cursor has not published the full model architecture or parameter count. The company describes Composer 2 as a mid-size model that prioritizes speed and accuracy on code tasks over general knowledge. This is a deliberate trade-off. You will not use Composer 2 to write marketing copy, but you will use it to restructure a React component tree across 15 files.
The Bigger Picture for AI Coding Tools
Cursor’s move signals a broader industry shift. Companies that started as wrappers around OpenAI or Anthropic APIs are now building their own models to reduce dependency and improve margins. GitHub Copilot took this path with its own fine-tuned models. Cursor is the latest to follow.
For developers, the result is better tooling at lower cost. For the AI industry, it raises questions about how long general-purpose model providers can sustain their current pricing when specialized alternatives keep arriving.
