A Cost-Efficient Revolution: Cursor Launches Composer 2.5, an AI Coding Powerhouse Rivaling GPT-5.5 in Performance

The field of AI-assisted programming has welcomed a formidable new disruptor. Cursor, the developer of the AI-powered code editor, recently officially launched a major upgrade to its self-developed programming large model—Composer 2.5. This model is built upon a fine-tuned version of Moonshot AI's open-source Kimi K2.5 base model. During training, the team utilized 25 times more synthetic task data compared to the previous generation.

To maximize the enhancement of the model's programming logic, Cursor allocated a staggering 85% of its computational budget directly to reinforcement learning and extended training. This aggressive R&D investment yielded an exceptionally significant performance leap, enabling the new model to demonstrate capabilities in handling complex coding tasks that rival those of the industry's top-tier models.

High Cost-Performance Ratio Challenges Industry Giants

In the widely recognized multi-language SWE-Bench test and the CursorBench v3.1 evaluation, Composer 2.5 achieved outstanding scores of 79.8% and 63.2%, respectively. This performance is on par with the most powerful models currently on the market, such as Anthropic's Opus 4.7 and OpenAI's GPT-5.5.

In stark contrast to its formidable performance is its remarkably low operating cost. The price for Composer 2.5 is only $0.50 per 1M input tokens and $2.50 per 1M output tokens, a mere fraction of the rates charged by competing large model developers.

Trillion-Scale Compute Forges the Next-Generation Heavyweight

For developers pursuing ultimate efficiency, Cursor also simultaneously released a faster variant. This version maintains equally excellent performance while costing only $3.00 per 1M input tokens and $15.00 per 1M output tokens.

This high-cost-performance AI programming service is now fully available for users to experience within the Cursor editor. Furthermore, Cursor has already partnered with SpaceX and xAI. Leveraging the colossal computing power of the Colossus-2 cluster, which hosts millions of H100 GPUs, they have begun training a next-generation, even more disruptive successor model from scratch using 10 times the computational resources.