Introducing Cursor Router

Announcements Release Discussions
, ,
1

Blog · Changelog

Today, we’re launching Cursor Router, our intelligent model router.

Not every request needs frontier-level intelligence. Cursor Router classifies each request by task type and complexity, then sends it to the most effective model for the job. Frontier models handle the work that demands them, and price-efficient models handle the rest.

In online A/B tests across millions of requests, it delivered near-frontier performance with significantly lower cost.

cursor-router-og

What’s new:

  • Select Auto in the model picker, and every agent request is analyzed and routed to the most cost-effective model that still delivers comparable quality.

  • Choose from three optimization modes, and switch between them anytime:

    • Intelligence, frontier quality on the hardest tasks.
    • Balance, the strong quality most people daily drive.
    • Cost, the previous Auto routing (with fixed per-token prices).

  • During early access, users saved roughly 31–52% on Auto-routed requests compared to routing everything to a single frontier model, with no drop-off in quality or keep rate.

  • The router can be enabled per-team or per-group. You can restrict which optimization modes members can select, set the default mode, allow or block underlying models, and choose whether the routed model is displayed or hidden. You can also make Auto the team default with soft or hard enforcement.

Cursor Router works across desktop, web, iOS, CLI, and the SDK, and is available today on Teams and Enterprise plans. It’s on by default for Teams, and Enterprise admins can enable it from the dashboard.

Read the docs here to learn more.

5

nice work！

6

Congratulations on the launch. The engineering quality is clearly high, and a whitepaper detailing the router observations, how it was designed and why would be a fascinating read.

My primary concern is the rollout communication.

Cursor’s “Auto”, to me, used to mean “cheap flat rate, don’t worry which model actually ran behind the scenes”. I emphasized used to because as of today, this no longer holds.

The now legacy “Auto” behavior seems to be “Auto Cost”?

If correct, I propose the new “Auto Cost” setting should ideally be aliased to the standard “Auto” that people are accustomed to.

This would prevent long-time users and newcomers alike from having to completely update their understanding of how the feature works - those who have been using Cursor for years now (like myself) or those just starting to use it, but also learning from Cursor veterans don’t have to suddenly change their mental model of “Auto”.

Historically, “Auto” meant a predictable, cheap flat rate regardless of the underlying model. Operating under that assumption today (July 22, 2026) is a significant financial risk.

It could also be my Cursor setup, but my “Auto” setting migrated to “Auto Balance” - which immediately started to draw frontier rates from my API tiers, which was very surprising and new behavior to me.

I will even be bold and say, this could end up with a lot of complaints! As of 2026-07-22 19:06 PST, I have not seen any banner, toast notifications, popups or emails notifying people that Cursor’s Auto behaves like this:

Balance and Intelligence now bills at whatever model the router actually picks and **draws from your API usage tier**.

So if you're still on "Auto" and see "GPT-5.6 Sol (Auto Balanced)" billed to your account under On-Demand, that's now **expected**: while you didn't pick Sol manually, Cursor's "Auto Balance" (which I believe is the default?) did, at Sol's API rates and billed you for it.

If I’m in the minority, surprised that “Auto” now bills me frontier rates, that’s fine. I just hope, I’m indeed in the minority.

Otherwise, I’m afraid, people would be excited about all this functionality being available under “Auto”, assume that, the now legacy “Auto” behavior, still applies, go “all out” and then be surprised at the invoice - it just doesn’t hold any longer for 2/3 Auto options.

7

awesomeeee

8

I have a feature request to be able to set Intelligence / Balanced / Cost as a default for Cloud Agents - thanks!

9

I have a question regarding prompt caching behavior when using Cursor Router in Auto mode: During multi-turn agent sessions where context (system prompts, tool definitions, attached files) remains largely the same across turns, does the router attempt sticky routing to keep turns on the same model/provider endpoint to leverage prompt caching (prefix caching)? If the model is switched dynamically, will this reduce the cache hit rate and incur additional cost?

10

Hey @daipl!

Good question. In our testing, even accounting for cache misses due to provider changes during a conversation, we still observed significant savings for users who choose Auto.

11

I see no options in the UI to do this. Basically there are 3 toggles for allowing cost, balance and intelligence and whether they get to see which model is used?

12

@DanTee This comes with Model Access Control, which is available on the Enterprise plan.

13

Don’t see this on latest update in the IDE?

14

@AndrewC Cursor Router is currently not available on individual plans!

15

Will Auto mode still consume first-party quota? When I have intelligence mode, what kind of model will the router choose?

17

Auto Cost draws from your first-party quota. Auto Intelligence and Auto Balance bill against whichever pool the selected underlying model belongs to, and you’ll be able to see which model that is!

18

It’s available only for Team plans, yet there was no mention of that in the announcement. Honestly, this is funny and shows how “professional” the Cursor team is.

And I’m sure that later it will become available for Individual plans only through MAX Mode, forcing users to enable MAX Mode just to use the Router …

What a mess this has become, guys.

20

@erraoudy This was mentioned in the blog, changelog, docs, and this post. I’m not sure how much more we could have done to communicate it!

21

Being able to control which models auto can select and seeing which model was selected by each individual chat have been a longstanding request for Auto:

That’s a great option for the Enterprise users. I’m on Teams, so even though Auto Intelligence and Auto Balance are available, I won’t be using them without having access to the above features.

Edit: We found how to enable the selected model being visible for the Teams account. So is it correct that only limiting the models available is limited to Enterprise?

22

Please give all users this options

23

maybe make it available to all users then

24

Mention it in the title, not the end of the post.

add “For Teams & Enterprise Plans!”

25

Is there any chance we could have the option to disable “fast” variants? I’m totally fine with auto choosing the model that is best, but I’m also completely fine waiting for the non-fast variant of the model to save on the api rate.

I hope you’ll consider it, thanks.