Composer 2

What is Composer 2?

Composer 2 is a coding model available in Cursor. It’s designed to help with coding tasks that may require multi-step reasoning and long action sequences, with the model trained and optimized specifically for agent-style coding workloads.

Cursor reports that Composer 2 delivers improved benchmark results (including Terminal-Bench 2.0 and SWE-bench Multilingual) compared with Composer 1, and attributes the gains to a continued pretraining run followed by reinforcement learning on long-horizon coding tasks.

Key Features

• Available in Cursor: Composer 2 can be used directly within Cursor’s environment for coding assistance.
• Improved coding benchmark performance: Cursor cites large improvements across the benchmarks it measures, including Terminal-Bench 2.0 and SWE-bench Multilingual.
• Reinforcement learning on long-horizon tasks: The model is trained through reinforcement learning on long action sequences, targeting tasks that require hundreds of actions.
• Multiple pricing modes (including a faster variant): Composer 2 has a standard option and a faster variant, each with different token pricing.
• Evaluation methodology and benchmark context: Cursor describes how its reported scores are computed (including use of Harbor for Terminal-Bench 2.0 and multiple iterations per model-agent pair), which helps clarify what “performance” refers to.

How to Use Composer 2

1. Open Cursor and select Composer 2 from the model options.
2. Start a coding task (for example, asking the assistant to implement or debug code).
3. If speed is a priority, switch to the faster variant mentioned in Cursor’s model details.
4. Review outputs and iterate as needed for your specific coding goal.

Use Cases

• Terminal-based coding workflows: Use Composer 2 when your task involves command-line or terminal-driven steps that can span many actions, aligning with Cursor’s focus on Terminal-Bench 2.0.
• Fixing or implementing changes on real codebases: For tasks that resemble SWE-bench Multilingual-style evaluation, use Composer 2 to work through multi-step coding changes.
• Long, multi-step development tasks: When a coding request is expected to involve many iterative actions (hundreds of steps, per Cursor’s description), Composer 2 is positioned for that pattern.
• Cost-conscious development: Choose the standard Composer 2 pricing mode when you want the cited input/output token rates; choose the faster variant when you want speed and accept the different token economics.

FAQ

What benchmarks does Cursor use to report Composer 2 performance?

Cursor mentions improvements on Terminal-Bench 2.0 and SWE-bench Multilingual. It also explains that its Cursor score for Terminal-Bench 2.0 was computed using the Harbor evaluation framework with default benchmark settings.

How much does Composer 2 cost?

Cursor lists $0.50/M input tokens and $2.50/M output tokens for Composer 2. It also lists a faster variant at $1.50/M input and $7.50/M output tokens.

Does Composer 2 support long action sequences?

Cursor states that Composer 2 is trained to solve challenging tasks requiring hundreds of actions.

Where can I try Composer 2?

Cursor says Composer 2 is available in Cursor, and it also references trying it in the early alpha of Cursor’s new interface.

Is Composer 2 the only model available in Cursor?

Cursor’s article discusses Composer 2 alongside other models, noting that for models besides Composer 2 it used the maximum of official leaderboard scores and scores recorded in Cursor infrastructure. The article does not specify all available models, but it clearly indicates a model menu in Cursor.

Alternatives

• Other coding models available in Cursor: If you need different speed/cost tradeoffs, switching to another model option within Cursor is a direct alternative, since Cursor discusses “other fast models” and provides a faster variant for Composer 2.
• Agent-oriented coding workflows without this specific model: If your workflow is mainly about terminal automation and iterative steps, consider alternative approaches such as using different model types that focus on terminal/agent evaluation patterns (without implying a specific brand).
• General-purpose LLM coding assistance: For simpler code generation or editing tasks, a general coding-capable assistant model may suffice, though Composer 2 is specifically framed as strong for long-horizon, multi-action tasks.