Cursor Router: интеллектуальный роутинг ИИ-моделей для команд с экономией до 60%

Cursor opened general access to CursorRouter on July 30, 2026 — an intelligent model router for teams and enterprise customers that, in online A/B tests across millions of requests, delivered frontier-model quality while saving up to 60%.

How CursorRouter works

CursorRouter is a classifier that analyzes each request before running the model and routes it to the most suitable model. It evaluates the request itself, the context, task complexity, and domain, matching this against what Cursor knows about each model's behavior.

The company describes the problem Router solves like this: about 60% of developers in Cursor pick one model as their everyday "workhorse." As a result, even simple tasks run at frontier prices, and AI spend grows faster than the quality of the results. Router changes this logic: routine work goes to the cheapest models, interface updates go to the model with the best "taste," and complex, long tasks go to frontier models with strong reasoning.

• General availability — July 30, 2026, for teams and enterprise customers
• Trained on 600,000+ live requests, optimized against a satisfaction metric (AFC)
• Three modes: Intelligence, Balance, Cost
• Cursor processes hundreds of millions of code requests per week
• Router accounts for cache misses both in training and in cost estimation

How much teams save

According to Cursor, in online A/B tests across millions of requests, CursorRouter delivered frontier-level quality at a 60% savings. During the early access period, dozens of companies got the same performance at 30–50% lower cost.

The company provides figures by mode. Auto Intelligence reaches the Fable model's level of result satisfaction at roughly 60% lower cost for teams, while also raising satisfaction by about 15% relative to Opus 4.8 at nearly the same price. Auto Balance comes out above Opus 4.8 in satisfaction at roughly 36% lower cost, and against GPT-5.6 Sol it delivers comparable quality at lower spend. The three modes — Intelligence, Balance, and Cost — let a team pick its own point on the "price–intelligence" curve.

Why online tests, not offline evals

Cursor deliberately measured the router's effectiveness with large-scale online A/B tests rather than offline evals. Offline tests, the company argues, are small in scale, disconnected from real usage, and don't account for the extra cost of cache misses when switching models.

"Our goal is to give teams the best performance and experience for every task, without spending more than the work requires,"

Cursor's blog states.

Real routing plays out inside a conversation: an engineer writes code, asks follow-up questions, catches errors, and keeps going — often hundreds of requests a week. Router has to pick the model and decide when to switch it under exactly these conditions. The classifier is designed for a world where new models ship often, so Cursor can quickly update the router as releases happen, without retraining the product from scratch.

What this means

Model routers are turning from an idea into a finished enterprise product: teams no longer need to pay frontier prices for routine work — cost now ties to the actual complexity of the task.

Frequently asked questions

Who can access CursorRouter?

CursorRouter has rolled out to general availability for Cursor's team and enterprise customers. It was previously available to dozens of companies as part of early access.

How much can you save with CursorRouter?

According to Cursor, in online A/B tests the router delivered frontier-model quality at a 60% savings, and during early access companies got the same performance at 30–50% lower cost.