Engineers from Axiom Math conducted the DrivingBench experiment between September 21 and 24, 2026, to test if large language models could operate a physical vehicle. The team used a 2022 Toyota Corolla equipped with open-source openpilot hardware from Comma.ai to translate software commands into vehicle actions.
OpenAI’s GPT-6 Astra was the only model to complete the 134.7-meter course, finishing in 5 minutes 22 seconds on its second attempt. The successful run consumed 6.6 million tokens, with a total cost of $9.75. The vehicle operated under constant human supervision in a low-speed, controlled environment.
Other models failed to finish the course. Anthropic’s Claude Fable 5.1 covered 45% of the distance on its third attempt, while xAI’s Grok 4.6 completed 11% on its best run. Engineers identified spatial perception issues as the primary cause for these failures.
Before the test, all models initially refused to operate the vehicle. They only agreed to proceed after the engineers labeled the task as a sandbox environment. The experiment gained traction online, receiving approximately 257 points on Hacker News.
