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

Cursor Router is the model routing system behind Auto, it has three modes rather than two, and Cursor documents it as available only on Teams and Enterprise plans. You reach it by opening the model picker, selecting Auto, and picking a mode under Optimize For: Cost, Balance, or Intelligence. (Cursor Router reference)

Key Takeaways

• Three modes, not two. Cost, Balance, and Intelligence, selected under Optimize For in the model picker after choosing Auto.
• Teams and Enterprise only. Cursor's reference page and its models-and-pricing page both say so.
• Enterprise has to switch it on. Cursor documents the router as off by default for Enterprise teams.
• Balance is the default for new users on plans that have the router.
• Cost mode bills as a Cursor-model request from the weekly pool; Balance and Intelligence bill at the routed model's list price, and a third-party route in those modes also incurs the Cursor Token Rate unless Cursor lists the model as excluded.
• Balance and Intelligence burn usage limits faster than Cost.
• Grok 4.6 is required in the pool. Blocking it disables the router entirely.
• GPT-5.5 is one of two recommended router models, and OpenAI plans to end its direct partnership with Cursor; see what OpenAI's Cursor cutoff means.
• The routed model is hidden by default, so you judge the answer rather than the model name.

How To Turn On Auto Balance or Auto Intelligence

Open the model picker in the chat or agent panel, select Auto, then pick a mode under Optimize For. Cursor says you can switch modes at any time. There is no separate model in the list called "Auto Balance" or "Auto Intelligence"; the mode is a setting on Auto. (Cursor Router reference)

Two conditions decide whether the option is there at all. Cursor's reference page states that "Cursor Router is currently only available on Teams and Enterprise plans," and its models-and-pricing page repeats it: "On Teams and Enterprise plans, Cursor Router picks the model for each Auto request based on your optimization mode." On Enterprise, an admin has to enable it first, because Cursor documents the router as off by default there. (Models and pricing)

The Three Optimization Modes

Cursor's own descriptions, quoted from its reference page:

Cursor adds one operational consequence worth knowing before you switch: "Balance and Intelligence use your usage limits faster than Cost." Its help center puts a number on the upside, as its own claim rather than an independent measurement: Intelligence "delivers about 20 to 30% higher quality" and is "recommended for complex, multi-step work." (Cursor Router help)

Which Models Cursor Router Routes Across

Cursor's available-models help page names the current pool: GPT-5.5, Claude Opus 5, Grok 4.6, and Claude Fable 5. The list is not decorative, because two entries carry hard requirements:

• Grok 4.6 is required. "The router needs a powerful, cost-efficient model in the pool; blocking Grok 4.6 disables the router." The reference page says the same: "enabling Grok 4.6 is a requirement for the router to work."
• GPT-5.5 and Claude Opus 5 are recommended. "Blocking one reduces routing quality. Blocking both disables the router."

On Enterprise plans the router respects team model-access controls and falls back to an allowlisted alternative when a model is blocked. Cursor is explicit that the pool is not stable: "the model pool changes over time as new models ship," and you cannot hand-pick the model for a request. (Available models)

GPT-5.5's place in that pool depends on a contract that is ending. Cursor staff wrote on August 29, 2026 that "OpenAI announced they plan to end their direct partnership with Cursor in three months," and that Cursor would "share more in the coming weeks around continuity of access to OpenAI models." (Cursor forum) Our coverage of OpenAI's Cursor cutoff has the proposed date and which Cursor surfaces lose access. If GPT-5.5 leaves the pool, Claude Opus 5 is the only recommended model left, and Cursor's rule is that blocking one of the two reduces routing quality.

What Cursor Router Costs

As of October 22, 2026, Cursor's reference page separates Cost from the other two modes: "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, except the models listed under Cursor Token Rate." The models-and-pricing page names the excluded models (as of the same date, 15 of them, including Claude Haiku 5.5, Gemini 3.8 Flash, GLM 5.3 Flash and GPT-5.6 Luna). Cursor's help center still carries the shorter line "All Auto modes bill at the routed model's list price"; the reference page is the more specific of the two. There is no routing surcharge and no flat Auto rate on standard plans, so a mode that reaches for frontier models more often simply spends more.

Enterprise had one dated exception, and it is over. Until September 7, 2026, Cursor's reference page set Enterprise Auto Cost pricing per million tokens regardless of the routed model. That sentence is no longer on the page, so Enterprise now follows the same list-price rule as every other plan with the router. Our coverage of Cursor's doubled usage pools has the surrounding pool structure, including why Auto no longer draws from the Cursor Models pool.

Admin Controls On Teams and Enterprise

Cursor documents four router settings in the team dashboard, and they change what individual members can do:

• Enable Cursor Router. Turns routing on or off for the team. Enterprise teams must enable it manually.
• Routing preferences. Chooses which optimization modes members can select. Up to two modes can be disabled.
• Underlying model. Displays or hides which model handled each response. Cursor says hidden is the default and the recommendation, "so results are judged on their own merit rather than by model name." It applies to Balance and Intelligence only.
• Impose Auto. Soft defaults each new chat to Auto while letting members switch; Hard locks the model picker to Auto. Cursor says both are off by default.

(Cursor Router reference)

How Model Selection Works

Cursor describes Router as a two-stage system. First, Compass, its complexity predictor, estimates whether a turn is simple enough for a price-efficient model, using a continuous complexity score from 0 to 1 and a threshold that decides when a request should be upgraded.

Compass learns from what users do next. Moving on to the next task is treated as a strong positive signal, while correcting the agent is treated as a strong negative signal. That lets the routing system learn from outcomes in real developer workflows rather than from a prebuilt benchmark alone.

For turns that need a frontier model, a second taxonomy classifies the work and identifies which model has shown the strongest observed performance for that kind of task. Cursor says the taxonomy covers domains such as backend, database schemas, and frontend; tasks such as fixing bugs, running commands, and writing tests; and modifiers such as bounded edits, product questions, and visual-heavy changes. (Cursor Router write-up)

Why The Model Can Change By Task

Cursor's argument is that no model dominates every kind of work. Its engineering write-up characterizes the models it routes across by first name only: Grok as "strong value across broad, routine work" with low inference cost, Sol as performing "especially well on planning and codebase comprehension," Opus as performing "well on execution-heavy work" including devops, database queries, and performance optimization, and Fable as excelling "at debugging and visual implementation."

The write-up does not attach version numbers to those names, so pair it with the routing pool Cursor documents today rather than assuming a mapping. The point is not that one model is always cheapest or always best; it is that the router can spend more of the budget where the task-specific quality difference is meaningful and keep routine turns on a lower-cost path.

What Cursor Says It Learned From Production Traffic

Cursor says its routing dataset contains hundreds of thousands of turns sampled from live Cursor traffic. Each datapoint includes the signals available to the router, a performance outcome, and a cost outcome. The company says it respected users' privacy-mode and data-retention settings while building the dataset.

The write-up reports that Compass's highest-confidence turns received a positive performance signal 96% of the time, while its lowest-confidence turns received one 71% of the time. Cursor also says it uses cross-validation, a held-out test set, and live-traffic testing to measure actual per-turn cost and user satisfaction.

Cursor launched Router on July 22, 2026 and, in its August 6 write-up, reported that the mode then called Auto Intelligence now delivers above Fable-level user satisfaction at 68% lower cost, a further 18% reduction since launch, and that Auto Balance outperforms Opus 4.8 at 41% lower cost, a further 8% reduction over the same period, while increasing user satisfaction by 3%. These are Cursor's own production measurements, not an independent benchmark, and they describe the modes as of August 6 rather than a guarantee about how they are tuned today. (Cursor Router write-up)

Using Router From The SDK

Cursor exposes Router in its TypeScript and Python SDKs as model id auto-smart, with a parameter named optimize_for that accepts cost, balanced, or intelligence. Note that the SDK value is balanced while the picker label is Balance. Cursor advises calling Cursor.models.list() first to confirm Router is available for the API key's team before hard-coding a selection, which matters given the Teams and Enterprise restriction. (Cursor Router reference)

Bottom Line

Cursor Router is a model-selection layer built around task complexity, observed developer outcomes, and a cost budget. Cost keeps the old Auto behavior, Balance is the default for new users, and Intelligence gives the router more room to spend on frontier models when Cursor expects the quality gain to matter. Before planning around any of it, check the plan: Cursor documents the router on Teams and Enterprise only, and off by default on Enterprise.

Corrections and Updates

October 22, 2026: This page said all three Auto modes bill the same way and quoted Cursor saying third-party models incur the Cursor Token Rate; Cursor's reference page now says Cost mode bills as a Cursor-model request from the weekly pool with no Cursor Token Rate, and the rate skips the models its pricing page lists as excluded.

October 9, 2026: This page said Grok 4.5 was the required model in Cursor Router's pool and that Enterprise Auto Cost had flat per-token pricing until September 7; Cursor now names Grok 4.6 as required, and that Enterprise pricing window has ended.

August 25, 2026: this page described two modes named "Auto Intelligence" and "Auto Balance" and never said which plans could use them.

Sources

• Cursor Router reference. Plan availability, the three Optimize For modes and their descriptions, usage-limit behavior, pricing, the Grok 4.6 requirement, admin controls, and the auto-smart SDK id. Re-checked October 9, 2026; the Enterprise Auto Cost sentence quoted on August 25, 2026 ("Until September 7, 2026, Enterprise Auto Cost pricing is set per million tokens") is no longer on the page.
• Cursor Router help center. "Legacy Auto is now called Cost in the Auto selection menu", Balance as the default for new users, and the 20 to 30% quality claim for Intelligence. Re-checked October 9, 2026.
• Cursor available models. The routing pool (GPT-5.5, Claude Opus 5, Grok 4.6, Claude Fable 5), the Grok 4.6 requirement, the GPT-5.5 and Claude Opus 5 recommendation, and hidden-by-default model identity. Re-checked October 9, 2026; on August 25, 2026 the pool and the requirement named Grok 4.5.
• Cursor models and pricing. "Auto has three modes: Cost, Balance, and Intelligence", and the Teams and Enterprise restriction stated a second time. Re-checked October 9, 2026.
• Cursor, "How Cursor Router chooses the right model for the task", August 6, 2026. Compass, the taxonomy, the dataset, the 96% and 71% confidence figures, the July 22 launch, and the cost and satisfaction figures. Re-checked August 25, 2026.
• Cursor Community, "I heard OpenAI is about to cut off GPT access for Cursor," staff posts of August 29 and September 7, 2026, read October 9, 2026: https://forum.cursor.com/t/i-heard-openai-is-about-to-cut-off-gpt-access-for-cursor-is-this-true/169876