OpenAI’s GPT-6 Astra has achieved what no commercial AI model has done before: it successfully drove a real car through a 134-meter course, completing the challenge in 5 minutes and 22 seconds at a cost of just $9.75 in API tokens.
The benchmark, called DrivingBench, was developed by researchers Aditya Ramabadran, Simon Mahns, and Tobias Gessler specifically to test whether general-purpose language models—rather than purpose-built autonomous driving systems—can handle the messy business of controlling a physical vehicle in the real world.
Three rival models attempted the same course and failed. GPT-6 Astra succeeded on its second try, navigating the complex real-world environment that includes unpredictable physics, sensor noise, and the need for real-time decision-making under uncertainty.
“This is a landmark result for the field,” the researchers noted. “It demonstrates that frontier models are approaching a level of physical world understanding that goes beyond text and code.”
The achievement raises the bar for what’s possible with general-purpose AI. While self-driving companies have spent billions on specialized autonomous systems, GPT-6 Astra accomplished the task using only its language model capabilities—interpreting sensor data, making driving decisions, and executing controls through an API interface.
OpenAI has not disclosed whether it plans to pursue autonomous driving applications, but the System Card for GPT-6 Astra, released earlier this month, outlined the model’s “Critical Threshold” capabilities for autonomous cyber operations. The driving benchmark suggests those capabilities extend to physical world tasks as well.
The $9.75 cost for a 134-meter drive also highlights the经济 efficiency of modern frontier models. At roughly 7 cents per meter, the experiment demonstrates that sophisticated AI control is becoming increasingly accessible for real-world robotics applications.
For the autonomous driving industry, the result is both inspiring and humbling—a reminder that the path to general AI may yield unexpected breakthroughs in places researchers weren’t primarily looking.