Cursor shipped Cursor Router on 22 July 2026, turning Auto mode from a single fallback into a routing layer that picks a model per request. It is a small release note with a large operational consequence: for the first time, the cost of an AI coding team is something an admin can tune rather than observe.
What actually shipped
The router exposes three optimisation modes, and the naming is unusually honest about the trade-off being made:
- Intelligence — routes for maximum quality, regardless of token cost.
- Balance — frontier-level models for standard day-to-day work.
- Cost — favours token-efficient models where the task allows it.
The part that matters for teams
The admin controls are the real story. According to Cursor's changelog, admins can enable the router per team or group, restrict which optimisation modes members can use, set the default mode, and allow or block specific underlying models.
That converts a per-developer preference into an organisational policy. A platform team can put the whole engineering org on Balance, allow Intelligence only for a named group, and block models that have not cleared review — without asking anyone to change how they work.
Why routing is showing up now
Routing appears when the model lineup gets wide enough that no single default is right. That is where 2026 has landed: the gap between a fast cheap tier and a top reasoning tier is now large in both directions — quality and price — and most requests genuinely do not need the top tier.
The honest reading is that routers are a margin tool as much as a quality tool. A team that was running everything through a frontier model was overpaying for autocomplete-shaped work. A router fixes that without a policy document.
What we would watch before switching everyone to Cost
Routing quality is only as good as the classifier deciding what a request needs. A misrouted refactor is more expensive than the tokens it saved, because a developer pays for it in review time.
The sane rollout is the boring one: put a small group on Balance, keep your hardest repository on Intelligence, and compare review churn over a couple of weeks before moving the whole org.