OpenAI’s GPT-6 Astra drives Toyota Corolla through parking lot course

The test, created by Aditya Ramabadran, Simon Mahns, and Tobias Gessler, required models to navigate cones using a comma four hardware setup with openpilot. Models had to interpret camera feeds, adjust steering and speed, and account for latency. Astra averaged under 1 mph, while Fable 5.1 reached 45% progress on its third try. Failures often stemmed from misreading lane markers or poor command timing. The researchers noted Astra showed in-context learning, adjusting its approach between attempts, while Grok and Sol struggled with perception and control cadence.

The benchmark is open-source, with all runs published as video traces. The team plans to expand it with more models, harder courses, and varied reasoning efforts. They emphasize the need for safer evaluation methods for real-world AI control.