GPT-6 Astra has gained the ability to drive a car

Generative AI models can now perform complex, real-world physical tasks like driving a car.

The DrivingBench project has demonstrated that GPT-6 Astra, an advanced AI model, can successfully drive a car. Researchers gave the AI control over a Toyota Corolla's steering, accelerator, and brakes to evaluate its performance on a cone course. In one attempt, GPT-6 Astra achieved 100% progress on the course within 5 minutes and 22 seconds, a significant improvement over other tested models like Claude Fable 5.1 and Grok 4.6. The model completed the course successfully after three attempts, with the best run showing 100% progress and a completion time of 5:22. Other models struggled, with many attempts resulting in a "Did Not Finish" status, indicating collisions or inability to navigate the course.

This development signals a significant leap in the capabilities of large language models, moving them from purely text-based or image-based tasks to direct physical control. The successful navigation of a car by GPT-6 Astra on a defined course suggests a potential pathway for AI to manage complex robotic operations and autonomous systems. It raises questions about the future integration of such models into safety-critical applications, requiring rigorous validation beyond simulated or controlled environments.

The performance difference highlights the rapid advancements in AI's ability to process sensor data, make real-time decisions, and execute precise physical commands. This could accelerate the development of more sophisticated autonomous driving systems and advanced robotics across various industries. Future research will likely focus on improving reliability, safety, and the ability of these models to handle unpredictable real-world conditions, potentially leading to new forms of human-machine interaction and automation.