Cursor released Cursor Router, a classifier that assigns each coding request to whichever AI model handles that specific task best, rather than locking an entire session to one model chosen at the start.
The system analyzes a request's query, context, task complexity, and domain, sending routine edits to cheaper, faster models and long, complex problems to frontier reasoning models. Early access customers with thousands of users cut AI spend by roughly 30 to 50 percent compared with routing every request to Anthropic's Opus 4.8, with no drop in output quality, according to Cursor's announcement; in broader online tests across millions of requests, the company reported savings as high as 60 percent.
Cursor built Cursor Router on more than 600,000 live coding requests, then checked the results with A/B tests across millions more, measuring how much AI-generated code developers actually keep in their codebase versus rewrite or reject. The tool ships with three modes, Intelligence, Balance, and Cost, that let teams set where they sit on the trade-off between raw model capability and token spend, and admins can restrict which models specific groups are allowed to use.
The launch responds to a pattern Cursor says is common among its own users: about 60 percent of developers pick one model as a default and stick with it, even when a cheaper or better-suited model would handle a given task just as well. Cursor Router is now available for Teams and Enterprise plans across desktop, web, iOS, CLI, and the company's SDK, positioning per-request routing as a lever for engineering teams watching AI spend outpace the code it actually produces.
