Cursor says its model router delivers frontier quality coding at 30% to 50% lower cost. That claim is doing a lot of work, and two details decide whether it applies to you.
The first is access. Cursor Router runs on Teams and Enterprise plans only, so if you are on an individual plan the feature is not available at any price. The second is billing. The three modes do not just pick different models, they draw from different budgets. Switching from Cost to Intelligence can lower your cost per task and raise your invoice at the same time, because the money comes out of a different pocket.
Here is what each mode does, what it actually charges, and a two week test that shows whether Router saves your team money or just moves it around.
What Cursor Router is, and who can use it
Cursor Router is the routing system behind Auto. Per Cursor's documentation, it runs a classifier on every agent request and sends it to a model matched to the task type and complexity: simpler requests go to fast, efficient models, harder ones to more capable models. Cursor manages the pool, and you cannot hand pick the model for a given request.
Cursor launched it on July 22, 2026. Three access details are easy to miss:
It is not available on individual plans. Free, Plus, Premium, Super and Ultra subscribers do not get it.
On Enterprise it is off by default. An admin has to enable it, and it can be configured per organization group.
Grok 4.6 must be enabled for the router to work at all. Cursor also warns that blocking too many models reduces routing quality and can disable the router outright.
The three modes compared
You choose a mode under "Optimize For" after selecting Auto, and you can switch at any time.
Mode | What it optimizes | How it bills | Burns usage limits |
|---|---|---|---|
Cost | Token spend, using the previous Auto routing logic | As a Cursor model request from your included weekly pool. No Cursor Token Rate. | Slowest |
Balance | Intelligence, speed and cost together | At the routed model's list price as third party usage, plus the Cursor Token Rate | Faster than Cost |
Intelligence | Sends harder tasks to the most capable models | Same as Balance: routed model list price plus the Cursor Token Rate | Fastest |
The billing detail that decides whether you save anything
This is the part most coverage of Cursor Router skips, and it matters more than the headline percentage.
Cost mode requests bill as Cursor model requests, drawn from the weekly pool your seat already includes. Balance and Intelligence requests bill as third party usage at the routed model's API list price, paid from prepaid credit, plus the Cursor Token Rate of $0.25 per million tokens on Teams and Enterprise. That rate applies to input, output and cached tokens alike. It does not apply to Cursor models such as Grok and Composer, to Cost mode requests, or to a list of smaller excluded models including Claude Haiku 5.5, Gemini 3.8 Flash and GPT-5.6 Luna.
What that means in practice. Moving a team from Cost to Intelligence does not just change model quality. It shifts spending out of the allowance you have already paid for and into metered credit. Your cost per completed task can fall exactly as Cursor describes while the amount you are invoiced goes up, because you were previously consuming included capacity and now you are buying tokens.
Run the arithmetic on the Token Rate alone
The Token Rate looks small and compounds quietly. The figures below are arithmetic from Cursor's published $0.25 per million rate, not measurements from a test we ran, and they exclude the model API charges that sit underneath.
Monthly tokens routed to eligible third party models | Cursor Token Rate charge |
|---|---|
100 million | $25 |
1 billion | $250 |
5 billion | $1,250 |
Before you enable Balance or Intelligence across a team, pull your current monthly token volume from the usage dashboard and multiply. If that number is material against your seat spend, it belongs in the comparison.
What Cursor claims, and what the claims do not cover
Cursor's launch announcement is unusually specific, which is to its credit. The figures below are the company's own, from online A/B tests rather than public benchmarks, and none have been independently verified.
Claim | Cursor's figure |
|---|---|
A/B tested savings at frontier quality | 60% savings |
Early access enterprises | Roughly 30% to 50% lower cost |
Cost per commit, Intelligence mode | $6.76, against $12.69 for Fable 5 |
Cost per commit, Balance mode | $4.63, against $7.34 for Opus 4.8 |
Training data for the router | More than 600,000 live requests |
Three gaps are worth holding in mind. The comparison baseline is routing every request to a single frontier model, which is not how most teams work, so your real saving is likely smaller than the headline. Cursor's quality metrics are user satisfaction and keep rate, which is how much agent written code survives in the codebase, rather than defect or rework rates. And cost per commit is a soft denominator: one developer's six small commits and another's single large commit can represent the same work.
Admin settings to review before you roll it out
Router is controlled from the team dashboard, and the defaults are opinionated. Walk through these before enabling it for everyone.
Setting | Default | Why it matters |
|---|---|---|
Show or hide the underlying model | Hidden | Cursor recommends hiding it so output is judged on merit. For a cost test you want it visible, or you cannot see what you are paying for. |
Disable optimization modes | All available | You can remove up to two modes. Useful to stop casual use of Intelligence on routine work. |
Impose Auto (Soft or Hard) | Both off | Soft makes Auto the default model, Hard locks the picker to Auto. Hard removes the escape hatch developers use when routing picks badly. |
Model access controls | Per organization | On Enterprise the router respects blocks and routes around them, but blocking too many models degrades routing or disables it. |
A note for regulated teams. Because you cannot choose the model per request, enabling Balance or Intelligence means your source code may reach any provider in the routed pool. If your organization restricts where code can be sent, configure model access controls first and confirm the remaining pool is large enough for routing to work.
A two week test you can actually run
Cursor's numbers describe Cursor's customers. This is how to get a number that describes yours.
Week 1, baseline. Keep your current setup. Record total AI spend, token volume, number of active developers, and the rough mix of work: routine edits, tests, debugging, larger features.
Turn on model visibility in the dashboard before week 2 so you can see which models requests land on.
Week 2, Router on. Same developers, similar work mix, one mode only. Start with Balance; testing two modes in one week gives you two half samples and no conclusion.
Log the rework. Ask developers to flag requests they repeated, answers they discarded, and code that needed substantial repair. Also log every time someone manually overrode the mode, otherwise you cannot tell whether Router performed or developers quietly rescued it.
What to measure
Track five things in both weeks. Put them in a spreadsheet with a column for each week.
Total AI charges, including prepaid credit drawn down, from the team dashboard's usage and billing view. A pass means this falls, not just that cost per request falls.
Charges per active developer, dividing by people who actually made requests rather than seats you pay for.
Repeated or discarded requests, tallied by developers on a sample rather than exhaustively.
Review time on agent written changes, from your PR tooling or by asking reviewers.
Cost per completed task, which is total charges divided by tasks shipped. This is the number that settles it.
Set the pass mark before you start. A reasonable one: total charges down at least 15%, with no meaningful rise in repeated requests or review time. Writing that down in advance stops the test turning into a search for a reason to keep the new setting.
Who should turn it on, and who should not
Router is worth testing if you run a Teams or Enterprise plan with varied work, enough volume that model spend is a line item someone asks about, and a mix that includes plenty of routine edits alongside hard problems. That mix is what routing exists to exploit.
It is less likely to help if nearly all your requests already need your strongest approved model, if compliance rules leave you with a thin pool of permitted models, or if your monthly spend is small enough that a 30% saving is not worth a workflow change. And if you are on an individual plan, the decision is made for you.
Frequently asked questions
Is Cursor Router available on individual plans?
No. Cursor's documentation lists it as available on Teams and Enterprise plans only. Individual subscribers still get Auto, but not the Balance and Intelligence optimization modes.
Can I see which model handled a request?
Only if an admin turns it on. Hiding the underlying model is the default, and Cursor recommends keeping it hidden so output is judged on quality rather than model name. Turn visibility on while you are measuring costs.
Does Cursor Router work through the SDK?
Yes. The TypeScript and Python SDKs expose it as model ID auto-smart with an optimize_for parameter accepting cost, balanced or intelligence. Cursor suggests calling the model list endpoint to confirm availability rather than hard coding it. The SDK runs agent workflows, not raw chat completions.
Will Router send my code to third party model providers?
In Balance and Intelligence modes it can, and you cannot pick the model per request. On Enterprise the router respects model access controls and routes around blocked models, so restrict the pool first if your code cannot leave certain providers. Note that blocking too many models can disable routing.
Does Cost mode use prepaid credit?
No. Cost mode requests bill as Cursor model requests from your included weekly pool and do not incur the Cursor Token Rate. Balance and Intelligence requests are billed as third party usage from prepaid credit.
The bottom line
Cursor Router is a sensible idea with a real mechanism behind it, and the company has published more detail about its evaluation than most vendors bother with. The catch is that its savings claim is measured against a baseline most teams do not actually run, and the mode you pick determines whether spending comes from capacity you already own or credit you have to buy.
Check your plan tier, check your token volume against the $0.25 per million rate, then run two weeks with the measures above. The number that matters is not Cursor's reported saving. It is what your team spends to finish code you would ship.