Introducing Cursor Router

Source: Cursor Blog — 2026-07-22

Summary

Cursor launched Router, an intelligent model-selection system for Teams and Enterprise plans that looks at each incoming request — the query, surrounding code context, task complexity, and domain — and automatically sends it to the most cost-effective model capable of handling it well. The classifier was trained on more than 600,000 live requests and validated in online A/B tests spanning millions of real requests, using user satisfaction as the reward signal. Early-access enterprise customers saw roughly 30-50% lower cost versus routing everything through a frontier model, with Cursor citing up to 60% savings in its broader A/B testing, and no drop in perceived output quality. Router ships with three selectable modes — Intelligence, Balance, and Cost — so teams can bias the tradeoff themselves.

Key Takeaways

• Router classifies requests before any model runs, weighing query content, surrounding context, task complexity, and domain, then combines that with learned data on how each candidate model actually performs on that type of task.
• Training data came from 600k+ live requests, with the resulting classifier stress-tested via online A/B experiments across millions of additional real requests, optimizing directly for user satisfaction rather than a static benchmark score.
• Reported savings: 30-50% lower cost among early-access enterprise customers compared to defaulting everything to a frontier model like Opus; Cursor's own A/B figures put maximum savings around 60% with no quality regression.
• The mental model is triage by task shape: routine edits and boilerplate go to cheap, fast models; UI/design-sensitive work goes to models with better "taste"; long-horizon, multi-step reasoning gets escalated to frontier models.
• Three user-facing modes — Intelligence, Balance, Cost — let teams set their own risk/spend tolerance instead of trusting the router's default judgment on every request.

Reel Script

This item has real training methodology and a measured cost-reduction metric — likely reel_eligible: true.

Hook Every time you fire off a one-line typo fix in Cursor, are you paying frontier-model prices for it? Most teams never check. They just eat the bill. Cursor just shipped something that stops burning your budget on tasks a cheap model could handle just as well.

Core Concept Cursor Router is basically a triage nurse for your prompts. Before any model actually touches your code, Router looks at the request, the surrounding context, how complex the task looks, and what domain it's in — is this a one-line fix, a gnarly refactor, a UI tweak — and decides which model should handle it. Think of it like a hospital ER: you don't send every patient to the head surgeon. A sprained ankle goes to a nurse practitioner, a heart attack goes to the specialist. Router does the same triage for your coding requests, except instead of a human making the call, it's a classifier that's learned, from real usage, which model actually performs best on which kind of task. Simple edits and boilerplate get routed to fast, cheap models. Work where visual or design judgment matters goes to models known for better "taste." And the deep, multi-step reasoning problems still get escalated to the frontier models — you're not losing capability, you're just not paying frontier prices for work that doesn't need it.

Hands-On Here's what makes this more than marketing spin: the classifier wasn't hand-tuned on a handful of examples. Cursor trained it on over 600,000 real, live requests pulled from actual usage, then ran it through online A/B testing across millions more real requests — not a synthetic benchmark, actual traffic — optimizing the whole thing for user satisfaction rather than some proxy metric. That's the part worth paying attention to: they measured whether people were happy with the output, not just whether it hit a score. The payoff Cursor is reporting: early-access enterprise customers saw costs drop somewhere between 30 and 50 percent compared to just routing everything through a top-tier model like Opus, and Cursor's broader A/B numbers claim savings as high as 60 percent with no measurable hit to quality. And critically, this isn't fully hands-off — Router ships with three modes: Intelligence, Balance, and Cost, so if you don't trust the algorithm's judgment on a given project, you can just tell it to bias toward quality or toward spend yourself.

Takeaway If you're running Cursor at team scale and still defaulting every request to your priciest model, you're leaving real money on the table. Router is a legitimate, data-backed way to cut spend without cutting output quality. Go check whether your team's on a plan that has it, and turn it on.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)