GPT-6 Astra navigates a Toyota Corolla in DrivingBench test

In the DrivingBench evaluation framework, OpenAI’s GPT-6 Astra became the first large language model to successfully navigate a physical vehicle through a pylon course. Using a 2022 Toyota Corolla equipped with comma.ai hardware, the model controlled steering, acceleration, and braking via text-based commands and MCP tools.

During the 134.7-meter course, Astra completed the track in 5 minutes and 22 seconds on its second attempt, moving at a speed of approximately one meter per second. While the model outperformed competitors such as Anthropic’s Claude Fable 5.1 and xAI’s Grok 4.6, researchers noted that the performance remains far from true autonomous driving. The experiment demonstrated that general-purpose models can interact with the physical world through coding environments, though they face challenges with spatial perception, latency, and visual data processing.

Safety measures included a speed limit of 3.5 meters per second and a human operator with access to an emergency brake. The project, developed by Aditya Ramabadran, Simon Mahns, and Tobias Gessler, highlights the potential for LLMs to manage chaotic physical environments using standard programming interfaces rather than specialized driving telemetry.