OpenAI's large language model GPT-6 Astra has become the first of its kind to successfully drive a car as part of the DrivingBench testing project. A full-fledged autopilot system is still a long way off: the car moved extremely slowly, and API access costs were high even with token caching enabled.
Experiment Course and Testing Results
During the tests, OpenAI GPT-5.6 Sol, SpaceXAI Grok 4.6, and Anthropic Claude Fable 5.1 failed to complete the test track entirely. Only OpenAI GPT-6 Astra managed to guide a Toyota Corolla along a 134.7-meter route on its second attempt, taking 5 minutes and 22 seconds to navigate a parking lot among cones.
Process Economics and Cost Comparison
The car covered the track at about 1.5 km/h, consuming 6.6 million tokens and spending $7.74 on API calls. Including hardware costs, the total operational expense reached $57.46 per kilometer. At standard fuel prices, using the AI model for driving turns out to be over 500 times more expensive than regular gas.
AI Perception Features and Expert Conclusions
Researchers admitted they deliberately tricked the models into believing the car was in a simulation, as real-world images caused panic in the neural networks. Experts believe long-term success lies in distilling powerful universal models into compact specialized versions.
