How Cursor Router Selects the Right Model for Each Task

Cursor Router learns how models perform in real development tasks to select the most suitable model for each request. The Auto Intelligence and Auto Balance modes, launched on July 22, have since been improved to reflect new models and operational traffic.

Currently, Auto Intelligence provides user satisfaction at or above the Fable level while reducing costs by 68%. Auto Balance cuts costs by 41% while outperforming Opus 4.8, with user satisfaction also increasing by 3% since launch.

Routing proceeds in two stages.

• Compass predicts whether the current request is simple enough to be handled by a cheaper model.
• More complex requests are routed to the optimal frontier model based on a task, domain, and modifier taxonomy learned from real development traffic.

The training data consists of hundreds of thousands of conversation turns collected from actual Cursor traffic. Each data point includes conversational signals available to the router, along with the following two evaluation metrics.

• Performance: Estimated by the user's next action; moving to the next task is considered a positive signal, while modifying the agent is considered a negative signal.
• Cost: Calculated using API prices and token usage, also reflecting cache miss costs incurred by model switching.

Compass predicts the likelihood of user satisfaction with the response in each turn, using this as a proxy for complexity. In actual online evaluations, requests assessed as having the highest success probability showed positive performance signals in 96% of cases, while even those assessed as having the lowest probability recorded positive signals in 71% of cases.

Compass assigns a continuous complexity score between 0 and 1. Requests below the set threshold remain with cost-efficient models, while those exceeding the threshold are upgraded to frontier models.