Scientists Download Frontier AI Model Into Self-Driving Car Let It Loose
TL;DR
In the DrivingBench project, three computer scientists connected frontier AI models, including OpenAI's GPT-6 Astra, xAI's Grok 4.6 and Anthropic's Claude Fable 5.1, to the controls of a Toyota Corolla. The car had to navigate a simple parking lot course marked with cones. Only GPT-6 Astra finished, on its second attempt, covering under 500 feet at 0.94 mph in five minutes, while the other models failed before the first corner.
Nauti's Take
The experiment shows real progress: general-purpose language models can steer a physical car without specialized training, albeit at walking pace. The bigger catch is the setup.
Persuading models to drive by telling them it is a simulation undermines exactly the safety guardrails that matter on real roads. For robotics teams this is a useful benchmark, while everyday autonomous driving still clearly belongs to specialized systems.
Sources