Cursor Router Launched for Teams and Enterprise Plans

Cursor announced the general availability of Cursor Router for its Teams and Enterprise plans this week. This new system functions as a classifier, analyzing each incoming request before it is processed by an AI model. It then directs the request to the model best equipped to handle that specific task.

Cursor reports that in online A/B tests, Cursor Router achieved frontier-quality performance while reducing costs by 60%. Early-access enterprise accounts have seen savings ranging from 30% to 50%. The company identified a significant spend pattern issue where developers often default to a single, high-cost model for all tasks, including routine ones. This leads to AI spending increasing faster than output quality improvements. Cursor Router is designed to address this mismatch by intelligently routing tasks.

The Cursor Router classifier is trained on over 600,000 live requests and evaluated across millions of live requests in online A/B tests. Its optimization goal is user satisfaction, measured by an "AI Feedback Cost" (AFC) reward signal. For each request, the router analyzes four key inputs: the query itself, the provided context, the complexity of the task, and the relevant domain. This analysis is combined with the system's learned knowledge of how different AI models perform.

Cursor has outlined three primary routing rules derived from its classification process. Simple tasks are directed to the most cost-efficient models. Requests related to UI updates are sent to models known for their superior "taste" or aesthetic judgment. Complex problems requiring long-horizon reasoning are routed to frontier reasoning models. Crucially, the cost savings are not achieved by downgrading the handling of difficult problems; instead, routine work is removed from the pricing tier of frontier models, while the capabilities for complex tasks remain accessible at their appropriate cost. The system is also cache-aware in both its training and evaluation phases, accounting for the cost of cache misses that can occur when switching models mid-conversation, a factor often overlooked by other routing mechanisms.