Here's how we built it:

Cursor Router chooses a model for each turn using two decisions. First, its Compass predictor estimates whether a cheaper model is likely to satisfy the user. More demanding turns then go through a task classifier that selects among stronger models using their observed performance on similar work. The technical explanation develops the mechanism behind the earlier Router launch.

Its training signal comes partly from what users do next: moving on is treated as positive, while correcting the agent is negative. Costs include the lost reuse of previously processed input when switching models. The system therefore balances observed user satisfaction against actual serving costs, including a cost that switching itself can introduce.