How Cursor Router chooses the right model for the task

This post details the technical implementation and data-driven approach behind Cursor Router's model selection. It introduces 'Compass,' a complexity predictor trained on user satisfaction signals (task completion vs. agent correction) to estimate turn complexity. The post explains how this complexity score is used to set a threshold for routing requests to either price-efficient or frontier models. It also describes the dataset construction from live traffic, capturing performance (user actions) and cost (API pricing, token usage, cache misses) as key metrics for training and evaluation. The core technical contribution is the system for learning optimal model routing based on real-world developer work, moving beyond benchmark scores.