Artificial intelligence programming is seeing a major disruptor. AI code editor developer Cursor has officially released a major upgrade to its self-developed programming large model - Composer 2.5. The model is fine-tuned based on the open-source Kimi K2.5 model from Moonshot AI, and during training, the team invested 25 times more synthetic task data than the previous version.

image.png

To maximize the model's programming logic, Cursor allocated 85% of its computing budget directly to reinforcement learning and extended training. This aggressive R&D investment has led to a significant performance leap, enabling the new model to demonstrate capabilities comparable to those of industry-leading models when handling complex coding tasks.

High Cost-Effectiveness Challenging Industry Giants

In industry-recognized SWE-Bench multi-language tests and CursorBench v3.1 evaluations, Composer 2.5 achieved excellent scores of 79.8% and 63.2%. This performance is already comparable to the most powerful models currently available, such as Anthropic Opus 4.7 and OpenAI GPT-5.5.

In stark contrast to its powerful performance is its extremely low operational cost. The price per million input and output tokens for Composer 2.5 is only $0.50 and $2.50 respectively, which is just a small fraction of the fees charged by competitors in the large model development sector.

Trillion-Level Computing Building the Next Generation Powerhouse

For developers pursuing ultimate efficiency, Cursor has also launched a faster variant version. This version maintains the same excellent performance while the price per million input and output tokens is only $3.00 and $15.00 respectively.

This high-cost-effective AI programming service is now fully available in the Cursor editor for user experience. In addition, Cursor has partnered with SpaceX and xAI to train the next generation of disruptive successor models on the Colossus-2 cluster, which has millions of H100 computing power, using ten times the computing capacity from scratch.