Cursor has officially released Composer 2, its latest AI model designed for advanced coding tasks. The company claims this iteration reaches "frontier-level" intelligence, representing a substantial leap forward in its capabilities. This release positions Composer 2 AI as a powerful new tool for developers seeking cutting-edge AI assistance.
The company detailed the performance gains in a blog post, showcasing significant improvements across key benchmarks. On Terminal-Bench 2.0, Composer 2 achieved a score of 61.3, a notable increase from its predecessor, Composer 1.5 (44.2) and Composer 1 (38.0). Similar gains were observed in SWE-bench Multilingual, where Composer 2 scored 73.7, outperforming previous versions.
Benchmark Breakthroughs
These performance enhancements are attributed to Cursor's initial continuous pretraining run, which built a stronger foundation for reinforcement learning. This approach allowed Composer 2 to be trained on complex, long-horizon coding tasks, enabling it to execute hundreds of actions to solve challenging problems.
The model's effectiveness on benchmarks like Terminal-Bench 2.0, which evaluates AI agents in terminal environments, highlights its practical application. Scores for models like Anthropic's Claude and OpenAI's models were also referenced, providing context for Composer 2's competitive standing.
Cost-Effective Intelligence
Cursor is pricing Composer 2 at $0.50 per million input tokens and $2.50 per million output tokens. A faster, albeit more expensive, variant is also available at $1.50 per million input and $7.50 per million output tokens, which Cursor states is cost-competitive with other high-speed models.