Based on Anysphere’s official docs, checked November 12, 2026

Cursor Auto mode and Cursor Router: Cost, Balance and Intelligence

Cursor Auto mode is no longer just a single default model choice. On Teams and Enterprise plans, Cursor’s documentation describes Cursor Router as the model routing system behind Auto: a classifier sends each agent request to a model that fits the task, so simple work can go to a faster, more efficient model and harder work can go to a more capable one. This guide follows Cursor’s Cursor Router docs, the Cursor Router help article and the available-models help article, all checked November 12, 2026. It explains the Cost, Balance and Intelligence modes, how Cursor says each is billed, the team admin controls, and how the TypeScript and Python SDKs expose Router as auto-smart. Cursor’s docs state that Cursor Router is currently only available on Teams and Enterprise plans. We have not run hands-on routing tests, and we do not rank models. The routing pool and the underlying models can change as Cursor ships and validates new models, so treat any named model list as the list on the page you checked, not a permanent lineup.

1. 1. Confirm you are on a plan that includes Cursor Router

   Cursor’s Cursor Router documentation says the feature is currently only available on Teams and Enterprise plans. If you are choosing Auto on another plan, do not assume you have the same Router modes or billing. On Enterprise, Cursor says teams must enable the router manually because it is off by default, and that the router can also be configured per organization group. A team admin should check the team dashboard before anyone treats Auto as Cursor Router. This page does not change your plan or billing; it only explains what the docs say.

2. 2. Select Auto, then pick an Optimize For mode

   Open the model picker, select Auto, and pick a mode under Optimize For. Cursor documents three modes. Cost uses the previous Auto routing logic and optimizes token spend. Balance optimizes for intelligence, speed and cost; the Cursor Router help article says Balance is the default mode for new users. Intelligence routes to the most capable models for harder tasks, at a lower cost than running a single frontier model. The help article also states that Intelligence “delivers about 20 to 30% higher quality by picking among first-party models in the pool.” That is Cursor’s claim, not an independent benchmark. Cursor says you can switch modes at any time, and that Balance and Intelligence use your usage limits faster than Cost.

3. 3. Know what you cannot control about the routed model

   Cursor says the router is data-driven and managed by Cursor. You cannot hand-pick which model handles a request. The model pool changes over time as new models ship. You steer routing by choosing an optimization mode, not by pinning a hidden model name. On Enterprise plans, the router respects your team’s model access controls: if a model is blocked, the router routes to an allowed model instead. Cursor warns that blocking too many models reduces routing quality and can disable the router. The Cursor Router docs say the router needs a powerful yet cost-efficient model for work that does not need other frontier models, and that enabling Grok 4.6 is a requirement for the router to work. The available-models help article, checked the same day, lists Grok 4.6 as required, and GPT-5.5 and Claude Opus 5 as recommended; it says blocking both recommended models disables the router. That pool description can change, so confirm it on Cursor’s page if you are setting an allowlist.

4. 4. Read the billing rules before you leave Cost mode

   Cursor’s pricing notes for Auto are specific, and they differ by mode. Cost mode bills as a Cursor-model request and draws from the weekly pool, with no Cursor Token Rate. Balance and Intelligence bill at the list price of the model each request is routed to. A third-party route in those modes also incurs the Cursor Token Rate unless the model’s cache-read price is strictly below $0.10 per million tokens across every pricing tier. The available-models help article states the same split for Teams and Enterprise with Cursor Router on. Because the router can choose a different model on the next request, a single chat can include more than one rate. If you need a reproducible comparison, Cursor’s SDK docs recommend a fixed model id rather than Auto. Do not invent a blended “Auto price”; look at the routed model and the mode you selected.

5. 5. Set team admin controls before you impose Auto on everyone

   Admins configure Cursor Router from the team dashboard. Cursor documents four controls. Enable Cursor Router turns routing on or off; when it is on, team members using Auto are routed by Cursor Router. Routing preferences let admins choose which optimization modes members can select, and Cursor says you can disable up to two modes. Underlying model shows which model Auto routed to at the start of each response, or keeps it hidden; hidden is the default and recommended, so results are judged on their own merit rather than by model name, and this setting applies to Balance and Intelligence. Impose Auto can make Auto the default for the team: a soft default starts each new chat on Auto but still lets members switch models, and a hard lock locks the model picker to Auto. Both Impose Auto options are off by default. Turn on a hard lock only if your team has already agreed that Auto is the right default, because it removes the ability to pin a known model for a sensitive task.

6. 6. Call Cursor Router from the SDK only after you discover it

   The TypeScript SDK and Python SDK expose Cursor Router as model id auto-smart, with a parameter named optimize_for. Cursor documents the values cost, balanced and intelligence. Call Cursor.models.list() to confirm Router is available for the API key’s team before you hard-code a selection. Cursor says Enterprise admins must enable Router for the team before auto-smart appears in the catalog. The SDK runs Cursor agent workflows; it is not a standalone chat-completions or raw inference API. Omitting optimize_for, or sending a value such as default, is not documented as a supported Router contract. If a specific model is missing from the catalog, Cursor also documents a server-selected Auto fallback as ModelSelection(id="auto"), and it tells you to prefer auto-smart when you need an explicit Router mode. Always discover allowed values rather than copying an example from an older post.

   import { Agent } from "@cursor/sdk";
   
   await using agent = await Agent.create({
     apiKey: process.env.CURSOR_API_KEY!,
     model: {
       id: "auto-smart",
       params: [{ id: "optimize_for", value: "balanced" }],
     },
     local: { cwd: process.cwd() },
   });

How this differs from picking one model

When you pick a named model, every request in that chat uses that model until you change it. Auto, with Cursor Router enabled, can send neighboring requests to different models. That is the point of routing, and it is also why Auto is a poor choice for a benchmark you want to compare later. Cursor says the underlying model can change between requests, and that you should prefer a fixed model id when you need reproducible comparisons.

It is also why hidden model names are the default in Balance and Intelligence. If your team needs to see the routed model for debugging or cost review, an admin can display it. Do not assume the model you saw on a previous turn is the model that will handle the next one.

Cursor Auto mode in this sense is not the same feature as Claude Code auto mode, which is a permission mode with a classifier that reviews actions. If you are looking for what runs without asking in Claude Code, use the Claude Code auto mode guide. The two “auto” names are easy to mix up in search results, and they are not interchangeable.

What the docs do not claim

Cursor does not publish a guaranteed quality score for Auto, and this page does not invent one. The 20 to 30 percent quality figure in the help article is Cursor’s own description of Intelligence mode, attributed here as such. Small differences on any vendor board may not be meaningful.

The docs also do not say that Auto is unlimited, cheaper than every named model, or available with the same Router behavior on every Cursor plan. If a help page mentions Auto being available in all regions, that is a geographic statement, not a plan-entitlement statement. For plan coverage, this guide stays with the Cursor Router documentation: currently Teams and Enterprise.

Finally, an allowlist that blocks Grok 4.6, or that blocks both recommended models named on the available-models page, can disable the router. If Auto suddenly disappears or degrades after a model-access change, check those requirements before assuming Router itself is down.

Tips

• Start in Cost if you want the previous Auto spend pattern; move to Balance or Intelligence only after you understand list-price billing.
• Use a named model, not Auto, when you need a reproducible comparison or a written record of which model produced a change.
• On Enterprise, confirm that an admin has enabled Cursor Router before you debug a missing auto-smart catalog entry.
• Do not block Grok 4.6 if you want the router to keep working; Cursor documents it as required.
• Claude Code auto mode is a different feature; do not copy its permission flags into Cursor.

Next steps

Sources

• https://cursor.com/docs/cursor-router · checked 2026-11-12
• https://cursor.com/help/models-and-usage/cursor-router · checked 2026-11-12
• https://cursor.com/help/models-and-usage/available-models · checked 2026-11-12
• https://prod.cursor.com/docs/sdk/python · checked 2026-11-12