Cursor Router: Request-Level Model Selection Cuts Costs 30-60% Without Quality Loss

Cursor introduced Cursor Router, a trained classifier that selects the optimal model per request, reducing costs 30–60% while maintaining frontier quality. It is trained on 600k+ real requests, optimized for user satisfaction, and uses four dimensions: query content, context, task complexity, and domain. The router is cache-aware, accounting for prompt cache invalidation costs. Three modes: Intelligence (near Fable quality, 60% cheaper), Balance (higher satisfaction than Opus 4.8, 36% cheaper). Enterprise early adopters saved 30–50% vs all-Opus. Evaluation uses online metrics (user satisfaction, keep rate) instead of offline evals. Dynamic tool calling further reduces token overhead.