DrivingBench puts frontier models in command of a real car's steering, throttle and brakes at walking pace. The gap between first and second place is the story.
Benchmark
DrivingBench
Whether a general-purpose language model can drive a real car through a fixed cone course in an empty car park, issuing one command at a time from what two cameras show it, with a human in the driver's seat ready to brake.
- Maintainer
- DrivingBench
- Score
- %
- Last checked
- September 23, 2026
Leaderboard
No verified results recorded yet.
Methodology
The model is given control of a comma-equipped Toyota Corolla's steering, accelerator and brakes through three tools — observe, set_motion and stop_now — and gets up to three attempts in one continuous chat session. Runs proceed at walking pace. The headline figure is progress: how far along the course centreline an attempt travelled while staying within four metres of it, as a share of the centreline length to the finish zone. Distance, finish time, accepted commands and token cost are recorded alongside it, with every command, the model's stated reason for it, the car's telemetry and the video aligned frame by frame.
Important limitations
A single low-speed cone course in a car park is not public-road driving, and nothing here tests traffic, pedestrians, weather or speed. Attempt counts are small, a human safety driver can end any run, and results are reported by the benchmark's own operators rather than by an independent auditor.