cursor shipped a model router so you stop overpaying for simple tasks
cursor router picks the right model for each request automatically — intelligence, balance, or cost mode — and claims 60% savings versus always reaching for the frontier model. it is live on teams and enterprise plans.
Cursor launched Cursor Routeron July 22, 2026 — an intelligent model router for teams and enterprise plans. Instead of sending every request to the same frontier model regardless of complexity, Router analyzes each request and routes it to the most capable model that still meets the task's requirements.
the performance numbers cursor is claiming are unusually specific: 30–50% lower cost in early access with dozens of enterprises, and 60% savings in online A/B tests across millions of requests, with "frontier-quality performance" on the output side.
three modes, one selection
cursor router exposes three optimization modes:
- intelligence — frontier quality, equivalent to always picking the most expensive and powerful model in the fleet
- balance — strong quality, matching the frontier models most developers use as their daily driver
- cost — good quality, highest available intelligence within a token-spend ceiling
router is available across desktop, web, iOS, CLI, and the SDK for teams and enterprise. it is not available on individual plans.
why this matters for vibe coding workflows
most agentic coding sessions are not uniformly hard. a long context refactor needs the frontier; renaming a variable or explaining a stack trace does not. routing on auto-pilot is a meaningful tax on a team budget when you multiply it across hundreds of daily requests per developer.
the bet cursor is making is that their classifier can make that call faster and more reliably than you can, session by session. the 60% savings claim needs a closer look — the A/B test conditions and benchmark definitions matter — but even half that rate would change the unit economics for a team running agents all day.
the honest test is to run balance mode for two weeks and compare your rejected diff rate against your old baseline. if quality held, you found your new default.
for teams already using agent sessions heavily, pair this with the four-prompt setup to make sure each task is scoped tightly enough that the router can make a clean call.