Cursor Composer 2.5 hits 79.8% on SWE-Bench

Cursor is proving that specialized post-training and massive synthetic data scaling can create a top-tier coding agent without the overhead of a general-purpose frontier model. New targeted RL with textual feedback allows for localized corrections of agent behaviors, while aggressive pricing at $0.50/$2.50 per million tokens puts pressure on competitors. The 79.8% SWE-Bench performance validates Cursor's synthetic training methodology, and the announcement hints at a future model being trained on the SpaceX-backed Colossus 2 cluster.