Composer 2 by Cursor

Composer 2: The New Frontier of Coding Intelligence in Cursor

In the rapidly evolving world of software development, Composer 2 has arrived as a groundbreaking update within the Cursor ecosystem. As a frontier-level coding model, Composer 2 represents an optimal combination of high-level intelligence and cost-effectiveness, designed to empower developers with advanced AI-driven programming capabilities.

Developed by the Cursor Team, this latest iteration sets a new standard for what AI coding agents can achieve, providing significant improvements in accuracy, task completion, and terminal proficiency.

What's Composer 2?

Composer 2 is a high-performance coding model integrated into the Cursor editor. It is specifically engineered to handle complex, long-horizon coding tasks that require hundreds of individual actions. Unlike standard language models, Composer 2 is the result of intensive research and development, including a first-of-its-kind continued pretraining run.

This foundation allows Composer 2 to scale effectively through reinforcement learning, making it one of the most capable models available for modern software engineering. Whether you are working within the traditional Cursor interface or experimenting with the new early alpha interface, Composer 2 acts as an intelligent partner that understands the nuances of multi-step development workflows.

Key Features of Composer 2

Frontier-Level Intelligence

Composer 2 delivers a massive leap in quality compared to its predecessors. By leveraging reinforcement learning, it can navigate challenging tasks that involve complex logic and extensive code modifications.

Benchmarked Performance

The superiority of Composer 2 is backed by data. It shows large improvements across all major metrics measured by the team:

• CursorBench: Scored 61.3, a significant rise from Composer 1.5's 44.2.
• Terminal-Bench 2.0: Achieved a score of 61.7, outperforming many other models in agent evaluation for terminal use.
• SWE-bench Multilingual: Reached a high of 73.7, demonstrating its versatility across different programming languages and environments.

Optimal Cost-to-Performance Ratio

One of the most compelling aspects of Composer 2 is its pricing structure. It offers frontier-level coding at a fraction of the cost of other high-end models:

• Standard Intelligence: $0.50/M input tokens and $2.50/M output tokens.
• Fast Variant: For those requiring lower latency, a faster version is available at $1.50/M input and $7.50/M output tokens.

Advanced Model Training

The success of Composer 2 stems from its unique training process. By using continued pretraining, the Cursor Team established a stronger base for scaling. The subsequent reinforcement learning focuses on solving tasks that require long-term planning and execution.

Use Case Scenarios for Composer 2

Composer 2 is designed for a variety of high-impact development scenarios where traditional AI might struggle.

1. Complex Refactoring

When a project requires changes across multiple files or deep architectural shifts, Composer 2 can handle the long-horizon actions necessary to ensure code consistency and functionality.

2. Terminal-Based Agent Tasks

Using the Terminal-Bench 2.0 framework as a guide, Composer 2 excels at executing commands and managing tasks within the terminal environment, acting as an autonomous agent to streamline developer workflows.

3. Multi-Language Software Engineering

Thanks to its high performance on SWE-bench Multilingual, developers working in diverse tech stacks can rely on Composer 2 to understand and generate code across various programming languages with high accuracy.

4. Rapid Prototyping

With the "Fast" variant as the default option, developers can iterate quickly, receiving high-quality code suggestions in real-time to move from idea to execution faster than ever.

Pricing and Usage

Composer 2 usage is designed to be accessible for both individual developers and larger teams:

• Individual Plans: Usage is part of a standalone usage pool that includes generous allotments.
• Fast Default: The faster variant, which maintains the same high intelligence level, is now the default option to ensure a smooth user experience.
• Transparency: Token sizing for Composer 2 is comparable to GPT models, with normalized pricing to ensure competitive rates against other fast models.

"Composer 2 delivers large improvements on all benchmarks we measure... making it a new, optimal combination of intelligence and cost." — Cursor Team

FAQ

Q: How does Composer 2 compare to Composer 1.5? A: Composer 2 shows significant improvements across the board. For example, in Terminal-Bench 2.0, it moved from a score of 47.9 to 61.7, and in SWE-bench Multilingual, it improved from 65.9 to 73.7.

Q: What is the cost of using Composer 2 in Cursor? A: The standard pricing is $0.50 per million input tokens and $2.50 per million output tokens. The fast variant is priced at $1.50/M input and $7.50/M output tokens.

Q: Is there a faster version of the model available? A: Yes, there is a faster variant with the same intelligence level. Cursor is making this fast version the default option for users.

Q: Can I use Composer 2 in the new interface? A: Yes, Composer 2 is available today in both the standard Cursor editor and in the early alpha of the new interface.

Q: What benchmarks were used to test Composer 2? A: The model was tested using CursorBench, Terminal-Bench 2.0 (an agent evaluation benchmark by the Laude Institute), and SWE-bench Multilingual.