Cursor has made Cursor Router generally available for Teams and Enterprise plans. The system is a request-level classifier that inspects each incoming request before any model executes, then dispatches it to the model best suited for that specific task. According to the Cursor team, online A/B tests demonstrate frontier-quality performance at approximately 60% cost savings, while three early-access enterprise accounts reported savings of 30–50%.

The underlying problem Cursor Router addresses is a spending pattern, not a capability gap. Cursor reports that roughly 60% of its developers use a single model as their daily driver. As a result, routine tasks are completed at frontier-tier prices, and AI expenditure grows faster than output quality improves. This tension is not unique to Cursor; across the AI coding assistant market — where Cursor competes with products such as GitHub Copilot, Codeium, and Tabnine — providers face the same challenge of balancing multi-model access against per-developer inference costs. Cursor Router is designed to resolve that mismatch.

How the Classifier Works

Cursor Router is neither a fallback chain nor a retry mechanism. It is a classifier trained on over 600,000 live requests, evaluated through an online A/B test spanning millions of live requests, and optimized for user satisfaction (AFC) as its reward signal.

For each request, the router analyzes four inputs: query, context, task complexity, and domain. These are combined with learned knowledge of each model's behavioral profile. Cursor publishes three routing rules derived from this classification:

• Simple work is routed to the most price-efficient models.
• UI updates are routed to the model with the best aesthetic judgment.
• Complex, long-horizon problems are routed to frontier reasoning models.

The third rule is central to the cost argument. Savings are not achieved by downgrading difficult problems; they come from removing routine work from frontier-level pricing while the hardest tier remains untouched.

A notable implementation detail: Cursor Router is cache-aware in both training and evaluation. It is trained on a dataset where routing produces cache misses, and the reported savings include the cost of those cache misses. Switching models mid-conversation invalidates prompt cache, and that cost is real — routers that ignore it overstate their savings. Prompt caching, which allows providers to reuse previously processed context at a discount, has become a standard cost-reduction feature offered by major model providers including Anthropic, OpenAI, and Google, making cache invalidation a materially expensive side effect of any model-switching system.

The classifier was also designed to accommodate rapid model churn. Cursor states the router can be updated as newer models ship, a meaningful advantage in a market where the frontier shifts on a monthly basis. With providers including Anthropic, OpenAI, Google, and xAI each releasing updated models on overlapping timelines, new options can appear before teams have finished evaluating the current ones.

Why Online A/B Tests Instead of Offline Evals

Cursor deliberately avoided offline evaluations as its primary measurement method. The company's stated reasoning is that offline evals suffer from small sample sizes, distance from real-world usage patterns, and the difficulty of reducing success to a standardized rubric. They also fail to account for the cache-miss cost incurred when switching models.

Real routing decisions happen across an entire conversation, not a single turn. Developers write code, ask follow-up questions, encounter errors, and continue — often across hundreds of requests in a given week. The router must determine both which model to select and when to switch between them.

Two quality metrics underpin the evaluation:

• User satisfaction: Agent success is classified from user responses. Moving on to the next feature is treated as a strong positive signal; correcting the agent is treated as a strong negative one.
• Keep rate: The proportion of agent-generated code that remains in the codebase over time.

The Cursor team states it has used both metrics to evaluate every model launch and harness improvement for the past nine months. These metrics predate the product they are now being used to validate. The keep rate metric reflects an emphasis on outcome-based measurement — whether AI-generated code persists — rather than task-completion rates that may not capture rework or downstream failures.

Three Modes and the Numbers Behind Them

Auto mode now exposes three optimization settings that move users along the cost–intelligence Pareto frontier:

• Auto Intelligence lands near Fable on user satisfaction at roughly 60% lower cost for teams. Against Opus 4.8, it lifts satisfaction approximately 15% at nearly the same cost.
• Auto Balance lands above Opus 4.8 on user satisfaction at approximately 36% lower cost. Against GPT-5.6 Sol, it delivers comparable satisfaction at a lower spend rate.
• Cost mode is described as achieving good quality while reaching the highest available intelligence and optimizing token spend. Cursor published no A/B quality or cost figures for this mode.

Since cost per request tells only part of the story, Cursor also measured cost per commit:

GPT-5.6 Sol matched the cost of Intelligence mode but produced lower user satisfaction. Cursor did not publish an exact per-commit figure for it.

Deployment and Procurement Constraints

Cursor Router is available across desktop, web, iOS, CLI, and the Cursor SDK. It is enabled by default for Teams plans. Enterprise administrators can activate it from the dashboard.

The changelog specifies the administrative surface: per-team and per-group enablement, restrictions on which optimization modes members can select, a configurable default mode, and model allow and block lists. Both soft and hard enforcement options exist for standardizing on Auto mode. The routed model can be displayed or hidden — it is hidden by default, so teams seeking routing transparency must explicitly opt in.

Two constraints are relevant for procurement planning:

1. Grok 4.5 is a required price-efficient routing option, meaning the model block list cannot be used to exclude it. Grok 4.5, released July 8, is priced at $2/M input and $6/M output tokens, with a fast variant at $4/M and $18/M.
2. Balance and Intelligence modes bill at the routed model's rate, so unit cost varies with each routing decision rather than settling at a flat per-request price.

Key Facts

• Cursor Router is a per-request classifier trained on 600,000+ live requests, optimized for user satisfaction (AFC).
• Online A/B tests report frontier-quality output at 60% savings; three early-access enterprise accounts saved 30–50% versus Opus 4.8.
• Cost per commit: $4.63 (Balance), $6.76 (Intelligence), versus $7.34 (Opus 4.8) and $12.69 (Fable 5).
• Cache-miss costs from model switching are included in reported savings, not excluded.
• Grok 4.5 is a mandatory routing option, and Balance and Intelligence bill at the routed model's rate.

Sources: Cursor Router launch post, Cursor Router changelog, Grok 4.5 announcement, and @cursor_ai on X.