OpenAI's GPT-6 Astra model successfully navigated a Toyota Corolla through a 134.7-meter test track, marking the first time a large language model has performed such a task. The vehicle moved at a speed of approximately 1.5 km/h, completing the route in 5 minutes and 22 seconds while navigating around cones.
The experiment, conducted as part of the DrivingBench project, required the model to process 6.6 million tokens, costing $7.74 in API fees. Researchers utilized a caching mechanism to reduce costs, achieving an effective rate of $1.17 per million tokens, significantly lower than standard pricing.
The total operational cost reached $57.46 per kilometer, including the use of a Comma Four driver assistance system and a smartphone. This figure is more than 500 times higher than the fuel cost for the same distance, which is estimated at $0.114 per kilometer based on local fuel prices.
Other models, including OpenAI GPT-5.6 Sol, SpaceXAI Grok 4.6, and Anthropic Claude Fable 5.1, failed to complete the course. To bypass safety restrictions, researchers informed the models that the vehicle was in a simulation, using the name DrivingBench Sandbox to maintain the illusion.
