Cursor Router picks the AI model per request

Cursor launched Router on July 22, a system that scores each request's complexity and domain before picking a model, per the company's changelog.

The stakes are practical. Picking the right model for every prompt across a team of engineers is a workflow tax nobody wants to own. Router removes that decision from the individual and hands it to policy.

Router runs on desktop, web, iOS, the CLI, and the SDK, and slots each request into one of three modes: Intelligence, Balance, or Cost.

Cursor says the routing hits frontier-level output at 60% less spend. That number comes from Cursor itself, and nobody outside the company has checked it.

What's new for Enterprise and Teams accounts is the admin panel underneath it. Admins can turn Router on per team or group, and set which mode is the default. They can also restrict which modes individual members are allowed to pick, or block specific models entirely.

That's the part worth watching. A 60% savings claim with no audit trail is marketing copy. A toggle that forces every engineer in a group onto Cost mode is something a platform team can flip. So is a rule that bans a specific model outright. What enterprises actually buy is the ability to enforce a policy across staff they don't fully trust to pick the right model themselves.