Cursor's Composer 2.5 Emerges as a Formidable Contender in AI Code Generation

Cursor has unveiled Composer 2.5, a small yet highly effective AI model specifically engineered for code tasks, which analysts suggest could be a significant, overlooked release. Priced at a competitive $0.50 per million input tokens and $2.50 per million output tokens, Composer 2.5 demonstrates performance comparable to leading models like GPT-5.5 and Anthropic’s Opus 4.7 on Cursor’s internal benchmark, “Cursor Bench.” While specific details of Cursor Bench are proprietary, the model achieved a 63% score, nearing GPT-5.5’s 64% and Opus 4.7’s 65%, but at a substantially lower cost. Developed from the open-source Kimmy K25 checkpoint, Composer 2.5 benefits from advanced training techniques, including targeted Reinforcement Learning (RL) with textual feedback to refine specific behaviors and extensive use of synthetic data—25 times more than its predecessor, Composer 2. This intensive post-training effort, backed by a collaboration with SpaceX AI providing substantial compute resources, has enabled Cursor to achieve rapid advancements, with internal tests indicating employees didn’t even notice a switch to Composer 2.5 for daily chat.

The release of Composer 2.5 signals Cursor’s strategic play in a competitive landscape dominated by major labs. By leveraging a substantial dataset derived from developer interactions within its platform, Cursor gains unique insights for training code-specific models. This approach allows Cursor to challenge the subsidization strategies employed by companies like OpenAI and Anthropic, which can make direct API usage by competitors financially prohibitive. However, a key limitation of Composer 2.5 is its exclusive availability within the Cursor ecosystem (app, CLI, and SDK), preventing independent external benchmarking via a public API. While this consolidates users within Cursor’s environment, it also raises challenges for broader integration and validation. Despite some observed UI performance issues with Cursor’s “Glass” interface during demonstrations, the model’s speed and efficiency make it particularly well-suited for highly collaborative development workflows within large enterprise codebases, an area where Cursor reportedly holds strong market penetration. Looking ahead, Cursor, in conjunction with SpaceX, is training an even larger model, “Colossus 2,” utilizing 100 times more compute than the original Kimmy base, hinting at a potential leapfrog in state-of-the-art code AI capability.