Powerful tools for Coding

Much stronger coding performance, with big benchmark gains over Composer 1.5 across CursorBench, Terminal-Bench 2.0, and SWE-bench Multilingual. Better long-horizon task solving, aimed at coding jobs that require hundreds of actions instead of shorter edit loops. Stronger base model training from Cursor’s first continued pretraining run, giving reinforcement learning a better foundation than before. Better price-to-intelligence balance, positioned as a much stronger model without giving up cost efficiency. Adds a faster same-intelligence variant and makes fast mode the default option.