How Cursor Router chooses the right model for the task · Cursor

Why it matters

Cursor has published a technical explanation of how its routing system selects which AI model handles each developer request. The system, Cursor Router, operates in two stages. First, a complexity predictor called Compass estimates whether a turn is simple enough for a price-efficient model, using a continuous score and a configurable threshold. Compass is trained on user behavior: moving on to the next task signals satisfaction, while corrections signal failure. Second, for demanding turns, a taxonomy built from real developer traffic classifies work by domain, task, and modifiers, then matches the turn to the frontier model with the strongest observed performance in that category. Cursor reports that no single model dominates all work; each of the frontier models it evaluated excels in different areas, such as planning, execution-heavy work, or debugging. The company says routing decisions are learned from live production traffic rather than benchmark scores, and that both routing configurations have improved since launch, with lower costs and, in one mode, higher user satisfaction.