GPT-6 Astra has gained the ability to drive a car
- Source
- drivingbench.com
- Author
- plurby
- Date
DrivingBench moves agent evaluation off the screen. Each model gets direct control of a real Toyota Corolla's steering, throttle and brakes and has to complete a cone course. GPT-6 Astra finished at 100% progress. Claude Fable 5.1 and Grok 4.6 both stalled well short of the end. The benchmark publishes full cost and token data for every run, so readers can compare what each attempt spent rather than only who finished. That detail is useful for teams building real-time or robotics agents, where latency and cost per decision matter as much as final success. A cone course at low speed is a narrow test and says little about general driving, so the result is better read as evidence about closed-loop control under time pressure than as a safety claim. It is still one of the few public benchmarks where a model's actions have physical consequences.
Context
OpenAI's GPT-6 Astra completed a real-car cone course on its second attempt in DrivingBench, according to the researchers' results, and the report says it was the only model to finish. The model received camera views and vehicle measurements, then issued steering and motion commands through software tools. Its successful run took 5 minutes 22 seconds.
The report makes the scope important: the course was in a parking lot, speeds were restricted, a human sat ready to brake, and each model's attempts shared one conversation so it could reflect on earlier failures. Each model was evaluated through one such sequence, and its attempts were not independent. This demonstrates a supervised physical-control capability in that setup; it does not measure dependable driving in public traffic.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Checking sign-in…
Loading comments…