Cursor Router slashes AI spend 60% by sending routine coding to cheaper models

Cursor has quietly solved one of the most expensive problems in AI-assisted development: paying frontier-model prices for routine work. Their new Cursor Router, now generally available for Teams and Enterprise plans, acts as a classifier that inspects every developer request before a model touches it, then dispatches it to the model best suited for that specific task. In online A/B tests across millions of live requests, the system delivered frontier-quality performance at 60% lower cost. Early enterprise accounts are seeing 30–50% savings.

The motivation is painfully familiar. Cursor found that roughly 60% of its developers pick a single model as their daily driver. That means simple completions — the kind of boilerplate and minor edits that dominate a coding session — get processed at the same premium rate as genuinely hard architectural problems. AI spend, in other words, grows faster than output quality. The router is Cursor’s structural answer to that mismatch.

Here’s how it works under the hood. This isn’t a fallback chain or a retry mechanism. It’s a classifier trained on over 600,000 live requests, optimized for user satisfaction as its reward signal. For each request, the router reads four signals — query, context, task complexity, and domain — and combines them with learned knowledge of each model’s behavior. Simple work goes to the most price-efficient models. UI updates go to the model with the best taste. Complex, long-horizon problems still get frontier reasoning models. The savings don’t come from downgrading hard problems. They come from removing routine work from frontier pricing.

One implementation detail separates serious routing from the hand-waving kind. Cursor Router is cache-aware in both training and evaluation. It’s trained on a dataset where routing produces cache misses, and the reported cost savings include the cost of those misses. Switching models mid-conversation invalidates the prompt cache — a real cost that most routers conveniently ignore. The company also designed the classifier for model churn, so it can update as newer models ship. That matters in a market where the frontier moves monthly. Three modes — Auto Intelligence, Auto Balance, and Cost — let teams slide along the cost-intelligence Pareto frontier. Auto Intelligence lands near Fable on user satisfaction at about 60% lower cost. Auto Balance beats Opus 4.8 on satisfaction while cutting spend 36%. The catch? Cursor published no quality or cost figures for Cost mode. The router ships across desktop, web, iOS, CLI, and the SDK, on by default for Teams, with admin controls for mode restrictions, model allow lists, and enforcement policies.