GPT-6 Astra Drives 2024 Toyota Corolla to In-N-Out, Completing 100% of Axiom's Test Course
Updated
Updated · WIRED · Oct 7
GPT-6 Astra Drives 2024 Toyota Corolla to In-N-Out, Completing 100% of Axiom's Test Course
3 articles · Updated · WIRED · Oct 7
Summary
Three Axiom engineers linked GPT-6 Astra to a 2024 Toyota Corolla’s cameras and steering, and the model drove the car through an In-N-Out drive-thru with a safety driver ready on the brake.
The test mattered because Astra is a general-purpose language model, not a driving system, suggesting newer multimodal AI can transfer 3D and visual reasoning into basic real-world control.
DrivingBench, the trio’s parking-lot benchmark, showed those skills remain limited: Astra was the only model to finish the course, while Claude Fable 5.1 reached 45% and Grok 11%, both slowly.
The engineers said the models appeared to improve in real time by adapting to the car’s controls, an emergent capability they tied to broader multimodal training rather than explicit driver training.
The stunt highlights physical reasoning as a key frontier for AI and robotics, with researchers arguing that stronger scene understanding could unlock home robots and other real-world applications while raising safety risks.
If a simple prompt trick can make an AI drive a real car, are our current safety guardrails fundamentally flawed?
How did engineers prevent a text-based AI from crashing a real car while it was still thinking about its next move?
Could general-purpose chatbots eventually replace billions of dollars of specialized self-driving technology by simply reasoning their way through traffic?