Cursor Router classifier cuts coding AI costs 30-50% while maintaining frontier quality

Cursor has released Cursor Router, a request-level classifier now generally available for Teams and Enterprise plans. The system inspects each incoming request and routes it to the model best suited for that task, delivering frontier-quality performance at 60% savings in online A/B tests and 30-50% savings for early-access enterprise accounts. Cursor reports that roughly 60% of its developers use a single model daily, causing routine work to be completed at frontier prices and inflating AI spend relative to output quality. The router addresses this mismatch by analyzing four inputs per request: query, context, task complexity, and domain. It was trained on 600k+ live requests and evaluated across millions of live requests in online A/B testing, optimized for user satisfaction as its reward signal. Cursor Router applies three routing rules: simple work goes to price-efficient models, UI updates go to the model with the best taste, and complex long-horizon problems go to frontier reasoning models. The classifier is cache-aware in both training and evaluation, accounting for the real cost of cache misses when switching models mid-conversation.