May 18th news, Cursor officially released its major upgraded version of the AI coding model developed independently - Composer2.5 today. This move marks a further enhancement of efficiency and cost-effectiveness in the field of AI programming.

This model is built upon the open-source Kimi K2.5 checkpoint from Moonshot (Moonshot Dark Side), with a drastic expansion in the scale of training tasks, reaching 25 times that of the previous product Composer2. It is reported that Cursor has allocated 85% of its computing resources to additional training and reinforcement learning (RL), thus achieving a breakthrough in core performance.

QQ20260519-091445.jpg

Composer2.5 performs similarly to Opus4.7 and GPT-5.5 on CursorBench3.1, but the cost per task is less than $1 - compared to the competitors' costs of up to $11. | Image: Cursor

In authoritative benchmark tests, Composer2.5 achieved an excellent score of 79.8% on SWE-Bench Multilingual, and a score rate of 63.2% on CursorBench v3.1. Its overall performance is now fully capable of competing with top flagship models such as Opus4.7 from Anthropic and GPT-5.5 from OpenAI.

While maintaining top-tier performance, Composer2.5 demonstrates high commercial competitiveness. The base version costs only $0.50 and $2.50 per million input and output tokens respectively, with a single-task cost of less than $1, far lower than competitors' charges of up to $11. Even for the high-performance fast versions, priced at $3.00 and $15.00 respectively, it still maintains an absolute price advantage.

Currently, Composer2.5 has been officially launched on the Cursor platform. Meanwhile, deeper strategic layouts have also been initiated, with Cursor deepening cooperation with SpaceX and xAI. Relying on the Colossus-2 cluster equivalent to the computing power of one million H100 chips, they are advancing the training of larger-scale subsequent models from scratch.