Cursor Router picks a model per request, cutting costs 30-50%
Cursor's new request-level router sends each coding task to the best-fit model, with early enterprise users reporting 30-50% savings and no quality drop.
Cursor released Cursor Router on July 22, an intelligent routing layer that classifies each incoming coding request and dispatches it to the model best suited for the job. The company says the system delivers frontier-quality results at up to 60% lower cost in online A/B tests, while early-access enterprise customers saw 30-50% savings versus sending everything to Claude Opus 4.8 — with no measurable drop in output quality.
How it works
Router is a request-level classifier trained on more than 600,000 live requests. It analyzes the query, context, task complexity, and domain, then applies a simple published logic: routine work goes to price-efficient models, UI changes go to models with stronger design instincts, and genuinely hard problems still go to frontier reasoning models. Cursor stresses that the savings come from pulling easy work off expensive models, not from downgrading difficult tasks. The classifier is also cache-aware — its training data and reported savings account for the cache misses that model-switching causes, an overhead many routing schemes quietly ignore.
Users pick between three modes — Intelligence, Balance, or Cost. In Cursor's published numbers, cost per commit runs $4.63 in Balance mode and $6.76 in Intelligence mode, against $7.34 for routing everything to Opus 4.8. Router is available today across all surfaces on Teams and Enterprise plans, where admins can enable it per team, allow or block specific models, and set defaults.
Why it matters
Model routing has long been pitched by startups, but Cursor shipping it as a default-on feature for teams signals the idea has matured from research demo to procurement line item. For enterprises staring at ballooning AI coding bills, a 30-50% cut with quality held constant reframes the buying question from "which model" to "which router." It also pressures the frontier labs: if a classifier can peel off the majority of requests to cheaper models, premium-tier pricing only holds for the hardest slice of work.