Cursor Launches Cursor Router: A Request Classifier to Cut AI Costs Up to 50% Without Quality Loss
Cursor
xAI
Anthropic
Cursor has released Cursor Router, a classifier that inspects each request before model execution and routes it to the best-suited model. In online A/B tests, the system achieves frontier-quality output at 60% cost savings, with early enterprise accounts reporting 30–50% savings versus Opus 4.8. The router is trained on 600k+ live requests and considers cache-miss costs in its savings calculations.
Cursor Router is now generally available for Teams and Enterprise plans. It is a per-request classifier trained on 600k+ live requests, optimized for user satisfaction (AFC). The system analyzes query, context, task complexity, and domain to route each request to the appropriate model. Cursor reports 60% savings in online A/B tests while maintaining frontier quality, and 30–50% savings for early-access enterprise accounts. The router is cache-aware, including cache-miss costs in reported savings. Cost per commit metrics show Auto Balance at $4.63, Auto Intelligence at $6.76, Opus 4.8 at $7.34, and Fable 5 at $12.69. Auto mode offers three optimization settings: Auto Intelligence (near Fable satisfaction at 60% lower cost), Auto Balance (above Opus 4.8 satisfaction at 36% lower cost), and Cost mode (optimizes token spend). Grok 4.5 is a required routing option, and Balance/Intelligence bill at the routed model's rate. The router ships across platforms and is on by default for Teams, with admin controls for per-team enablement and model restrictions.
- Сокращения
- AFC = Agent Feedback Classifier — Классификатор обратной связи агента
- CLI = Command Line Interface — Интерфейс командной строки
- SDK = Software Development Kit — Комплект разработчика
Source: MarkTechPost —
original