It's a bit harsh The new Composer 2.5 model from Cursor is here, directly benchmarking against Opus 4.7 and GPT-5.5. With a million input tokens, it only costs $0.5, and the usage limit will double next week! Cursor announces the release of the latest Composer 2.5 model. The model has achieved significant improvements in reliability in handling long-term tasks and complex instructions, and Cursor has also announced that it will provide users with double the inclusion usage next week. In benchmark testing, Composer 2.5 scored 69.3% and 79.8% on metrics such as Terminal Punch 2.0 and SWE Bench Multilingual, respectively, which are close to the levels of Opus 4.7 and GPT-5.5. However, the input cost of the standard version API has been significantly reduced to $0.5 per million tokens. Composer 2.5 is not built entirely from scratch, but is developed based on the latest 1 trillion parameter multimodal open source model Kimi K2.5 checkpoint released by Moonshot. The cursor team has focused 85% of its computing resources on additional reinforcement learning (RL) and fine-tuning, by introducing targeted text feedback mechanisms, 25 times new synthesis tasks, and optimized distributed training methods, with a focus on reducing the tool call error rate of the model in ultra long contexts and improving the accuracy of code modifications.