Researchers put frontier AI models in control of a Toyota Corolla's steering, accelerator, and brakes to test real-world driving capability. The evaluation took place on a fixed cone course where models had to navigate while staying within 4 meters of the centerline. Performance was measured by attemptprogress—the distance traveled along the course centerline before any collision—and finish time for completed runs.

Metrics also tracked total tokens and cost at list prices for each attempt. The setup, named DrivingBench, allows followers to track progress via @drivingbench on social platforms. Collision events reset progress to the furthest point reached prior to impact.

GPS speed data helped calculate integrated distance traveled during each drive. Tests included both successful runs and failures marked as DNF (did not finish).