Cursor Router: which mode to pick for your team (July 2026)

Cursor switched Router on for every Teams plan today, and the only decision it hands you is which optimization mode your seats run. Solo dev on Pro: leave it on Balance and get back to work. Team lead: keep Balance as the org default, then let the two or three people who live in agents all day opt into Intelligence. Enterprise admin: enable Router per group, set Balance as the floor, and lock your model allowlist before anyone routes a single request. Everything below is the reasoning behind those three calls.

We wired the admin side up this morning against the July 22, 2026 changelog, so the mode names, billing rules, and dashboard controls here match what shipped, not what a beta post promised.

What Router actually is, and how it's different from Auto

Router analyzes each request and sends it to the model best suited for that job, per Cursor's July 22, 2026 changelog. A frontier model handles the request that genuinely needs one. A cheaper, faster model handles the request that doesn't. You stop hand-picking a model per task, and you stop paying frontier rates for a one-line completion.

If that sounds like the old Auto mode, here's the distinction that matters. Auto made one opaque pick and mostly kept you inside Cursor's first-party model pool. Router makes the same per-request decision, but you tell it what to optimize for. Three named targets. It also runs everywhere the editor does: desktop, web, iOS, the CLI, and the SDK, per the changelog. And it comes with admin governance that Auto never had, which is the part that changes how a team manages spend.

The catch is that "let it decide" is only half the feature. The other half is deciding what "best" means for your team, and that is the mode.

The three modes, and which task goes where

Cursor ships three optimization modes. The descriptions below are Cursor's own, from the July 22 changelog:

• Intelligence targets "frontier quality, matching the most expensive and powerful models that might be out of reach for daily use."
• Balance targets "strong quality, matching the frontier models that most people like to daily drive."
• Cost targets "good quality, reaching the highest available intelligence while optimizing token spend."

Here's how that maps to what you'd actually run each one on.

Mode	What Cursor optimizes for	How it bills	Run it on
Intelligence	Top-tier model quality, cost secondary	The routed model's own rate, per changelog	A gnarly cross-service refactor, a hard debugging session, architecture reasoning where a wrong answer costs an hour
Balance	The frontier models people daily-drive	The routed model's own rate, per changelog	Everyday Composer edits, feature work, PR review, the bulk of agent runs
Cost	Highest intelligence it can reach while trimming token spend	Optimized for token spend, per changelog	CI-integrated review, high-volume SDK scripts, batch codemods, boilerplate generation

Balance is the one most teams should default to, and it's what the name implies: the models a working developer would reach for anyway, chosen per request so you're not overpaying on the easy ones. Intelligence is not "always better." It's "always expensive," and it only pays off on the class of task where model quality is the bottleneck. Cost is the sleeper. For anything you run at volume in a pipeline, where a good-enough answer at a fraction of the tokens beats a frontier answer you didn't need, Cost is the correct setting and most teams forget it exists.

Setting it up as a team admin

Router is on by default for Teams plans, per the changelog. Enterprise admins flip it on from the dashboard. Either way, the governance controls are where the real setup happens, because the defaults are permissive and your finance team will notice.

From the admin dashboard you can enable Router per team or per group, restrict which optimization modes members are allowed to use, set the default mode, and allow or block specific underlying models, all per the July 22 changelog. That last one is an allowlist and a blocklist, and it's the control most teams will lean on hardest.

Here's the rollout we'd run on day one:

1. Dashboard  ->  Router  ->  enable for the whole org
2. Set the default mode to Balance for every group
3. Restrict members to Balance + Cost only
   (leave Intelligence unchecked at the org level)
4. Create a "Platform" or "Senior" group  ->  grant Intelligence
5. Set the model allowlist BEFORE inviting the group in

The reason to restrict Intelligence at the org level and grant it per group is simple: an unrestricted Intelligence mode is an invitation for every seat to route every request to the most expensive model available, whether the task needs it or not. Grant it to the people whose work justifies it. Everyone else gets Balance and Cost, which covers the daily loop without a surprise line item.

The cost math, and how it hits your usage pools

This is where Router stops being a convenience and starts being a budget lever.

Balance and Intelligence bill at the routed model's rate, per the changelog. There's no flat Router surcharge on top. You pay for whatever model it picked, which means Intelligence mode is only as expensive as the frontier models it keeps reaching for. Cost mode is the one built to trim token spend while still reaching the highest intelligence it can afford.

Now layer that onto how Teams billing already works. Every Teams seat carries two separate pools of included usage: a Composer and Auto pool reserved for Cursor's first-party models, and a third-party API pool for frontier models like Claude, GPT, and Gemini, per our Cursor pricing guide. Router's mode is effectively a dial on which pool drains faster.

Push a seat to Intelligence and it routes toward third-party frontier models more often, which draws down that third-party API pool and tips you into on-demand billing sooner. Set it to Cost and more work stays on the cheaper models, so the included pools last longer. Balance sits in the middle by design. The mode you assign is, in practice, a prediction about which pool a seat will exhaust first.

The seat tiers make this concrete. A Teams Standard seat runs $40/user/month ($32 on annual), and a Premium seat runs $120/user/month ($96 on annual) with roughly five times the included usage, per Cursor's Teams pricing post and our own pricing guide. A Standard seat on Intelligence is the fast path to on-demand overages. A Premium seat on Intelligence is what that extra pool is for. So the mode assignment and the seat assignment are the same decision looked at from two angles, and treating them separately is how teams overspend.

The SpaceX model question

There's a governance reason to care about the model allowlist beyond cost, and it's the ownership question hanging over Cursor.

SpaceX agreed to acquire Anysphere, the company behind Cursor, for $60 billion in all-stock, announced June 16 with the deal expected to close in Q3 2026, per our Cursor review. SpaceX had already merged with xAI earlier in the year, so the eventual owner is effectively the Grok stack. Access to Claude and GPT is confirmed through close. The post-close model lineup is not announced.

Router's allowlist is the control that lets you not wait to find out. If your team has a policy about which vendors' models can see your source, or you simply want to guarantee that a future default doesn't quietly start routing your code to a first-party Grok model, you set the allowlist now and the router can only pick from what you permit. Block the vendors you won't ship code through; allow the ones you've cleared. It's a per-team and per-group control, so a regulated group can run a tighter allowlist than the rest of the org.

For a solo dev this is a non-issue today. For an enterprise standardizing on Cursor across hundreds of seats, the allowlist turns "we'll deal with the acquisition later" into a setting you configure once and audit against.

Which mode to pick

Balance for the org default. It's the setting Cursor built for the developers who daily-drive frontier models, it bills at the routed model's actual rate with no surcharge, and it covers the overwhelming majority of Composer edits, feature work, and agent runs without reaching for the priciest tier on tasks that don't need it.

Intelligence only for the agent-heavy seats that already justify Premium. If a developer lives in long agent sessions and multi-file refactors all day, they are the person whose Premium pool exists to absorb frontier usage, and they're the one whose output quality actually moves when the model gets smarter. Grant Intelligence to that group, not to the org.

Cost for anything scripted or high-volume. CI-integrated review, SDK batch jobs, codemods, and boilerplate generation don't need a frontier model, and Cost mode is the one tuned to reach usable quality while trimming token spend. This is the mode most teams underuse, and it's free money left on the table for any pipeline you run at scale.

The one setup step nobody should skip: lock the model allowlist before the acquisition closes, per our Cursor review. It's a two-minute dashboard change today and a governance headache if you wait.

If your team isn't on a Cursor plan yet, start on Cursor's pricing page to see the current Teams tiers, then work through the seat math in our Cursor pricing guide before you assign a single mode. For the full picture on the editor itself, the Tab model, the pricing complexity, and the ownership risk, our Cursor review has the decision framework.

See Cursor's current Teams pricing