Three engineers connected OpenAI’s GPT-6 Astra to cameras and the steering system of a 2024 Toyota Corolla and used it to navigate an In-N-Out drive-through. A human safety driver kept a foot over the brake throughout the experiment.
Unlike a conventional autonomous-driving stack, Astra is a general-purpose multimodal model rather than software trained specifically for vehicles. The engineers prompted it to interpret camera views and issue steering commands through a server. They say the model adjusted to control errors during the run despite receiving no dedicated driving training.
A separate parking-lot benchmark created by the group shows how limited the capability remains. Astra completed the simple course, but slowly. Anthropic’s Claude Fable 5.1 covered 45 percent and xAI’s Grok 11 percent. The models initially refused to control a real vehicle and were persuaded through prompting, which also raises questions about the reliability of their safety boundaries.
The demonstration is evidence of rudimentary physical reasoning, not a road-ready autonomous-driving system. It involved a controlled route, a modified car and immediate human backup. General-purpose models still lack the validation, redundancy and predictable failure handling required for safety-critical driving.