Composer 2 Launches with Frontier-Level Coding Intelligence and Cost-Efficient Pricing

Composer 2 is now live in Cursor, bringing a significant leap in coding intelligence while maintaining a cost-effective pricing model. Priced at $0.50 per million input tokens and $2.50 per million output tokens, it offers an optimal balance of performance and affordability.

Frontier-Level Coding Intelligence

The latest version, Composer 2, showcases substantial improvements across various benchmarks, solidifying its position as a top-tier coding model. Here’s how it stacks up against previous versions:

• CursorBench: 61.3% (up from 44.2% in Composer 1.5 and 38.0% in Composer 1)
• Terminal-Bench 2.0: 61.7% (up from 47.9% in Composer 1.5 and 40.0% in Composer 1)
• SWE-bench Multilingual: 73.7% (up from 65.9% in Composer 1.5 and 56.9% in Composer 1)

These improvements are driven by our first continued pretraining run, which provides a stronger foundation for scaling reinforcement learning. This approach allows Composer 2 to excel in long-horizon coding tasks that require hundreds of actions.

Technical Details

• Continued Pretraining: The initial pretraining phase is crucial for building a robust base model. From there, we fine-tune the model using reinforcement learning (RL) on complex, multi-step coding tasks.
• Reinforcement Learning: We use RL to train the model on long-horizon tasks, which are essential for real-world applications where multiple steps of reasoning and execution are required.

Pricing and Performance

Composer 2 is available in two variants:

• Standard Variant:

  • Pricing: $0.50 per million input tokens, $2.50 per million output tokens
  • Performance: Balanced for a wide range of tasks

• Faster Variant:

  • Pricing: $1.50 per million input tokens, $7.50 per million output tokens
  • Performance: Offers the same intelligence as the standard variant but with faster execution times, making it more cost-effective than other fast models.

Usage and Availability

• Individual Plans: Composer usage is part of a standalone usage pool included in individual plans. This ensures generous usage limits for all users.
• New Interface: Try Composer 2 today in Cursor or in the early alpha of our new interface, Glass.

For detailed information on model capabilities and pricing, refer to our model documentation.

Benchmarks

• Terminal-Bench 2.0: This benchmark evaluates agents for terminal use and is maintained by the Laude Institute. Our scores were computed using the official Harbor evaluation framework with default settings.
• SWE-bench Multilingual: This benchmark assesses performance across multiple programming languages, making it a comprehensive measure of coding intelligence.

Conclusion

Composer 2 represents a significant step forward in AI tooling for developers. With its frontier-level coding intelligence and cost-effective pricing, it’s an excellent choice for both individual developers and teams looking to enhance their productivity.