Cursor ships a router to cut model costs 60%

Cursor launched Cursor Router on July 22, a request-level classifier that reads each coding request’s query, context, task complexity, and domain, then routes it to the model Cursor judges best suited to the job rather than sending everything to one default model. Cursor trained the classifier on more than 600,000 live requests and says it tested it across millions more. The company claims frontier-quality results at up to 60% lower cost versus routing everything to a single top-tier model.

The router ships with three modes on a cost-intelligence tradeoff: Intelligence, for difficult diagnosis, architecture, and long-horizon work where a wrong answer is expensive to unwind; Balance, the default for typical day-to-day product work; and Cost, for bounded, repeatable tasks where minimizing spend per accepted task matters most. Cursor’s own numbers put Balance mode’s satisfaction score above Opus 4.8 at roughly 36% lower cost per commit. It’s available now on desktop, web, iOS, CLI, and the Cursor SDK, on by default for Teams plans, with admin controls for Enterprise to restrict which modes teams can use.

This is the same routing pattern spreading across the industry this week: Runway also launched a model router for generative media, and both follow the same logic, that a good classifier picking the cheapest model that still clears the bar beats manually picking one model for everything. Cursor’s own benchmark is self-reported, so treat the 60% figure as a ceiling under ideal conditions rather than what you’ll see on your specific codebase, and validate Balance mode against whatever you use as your default today before rolling it out team-wide.