GPT-6 Astra is the first artificial intelligence model to drive a real car.

GPT-6 Astra, the artificial intelligence model developed by OpenAI, has become the first general AI system capable of driving a real car. It demonstrated this by passing the DrivingBench test, an independent benchmark that put four different language models behind the wheel of a Toyota Corolla. Among them, only GPT-6 Astra successfully completed the trial, crossing the finish line.

The test represents a milestone in the application of general artificial intelligence to complex, real-world physical tasks, such as driving a vehicle. While the other three language models subjected to the same test failed to complete the task, GPT-6 Astra demonstrated integrated capabilities in perception, decision-making, and motor control.

The experiment, conducted on September 25, 2026, marks a qualitative leap compared to specialized autonomous driving systems, as it uses a general language model not specifically designed for driving. Operational details on how the model interacted with the car's controls and interpreted the environment have not yet been disclosed.

Autonomous driving is a field of research that has existed for decades, with the first experiments dating back to the 1980s. Traditional systems, like those from Tesla or Waymo, are "narrow AI": specialized software, programmed with millions of lines of code and trained on massive datasets of images and driving scenarios to recognize objects, follow signals, and make decisions. These systems are built for a specific task: driving a car.

The arrival of Large Language Models (LLMs), like OpenAI's GPT, has shifted the paradigm. These are "artificial general intelligence" (AGI) models, trained on the world's text and images to understand and generate language, solve problems, and reason. The challenge has been making them interact with the physical world. The project began when researchers wondered if a model, capable of understanding a driver's manual and describing a road, could also control a car.

The decisive step was the development of "embodiment," meaning the model's ability to translate its reasoning into physical actions through an interface (in this case, the car's controls). GPT-6 Astra represents the evolution of this path: it is not an autonomous driving software, but a general language model that has been given the ability to act. The DrivingBench test was the first real-world trial to verify this capability. Its success paves the way for future applications where a single general AI will be able to perform multiple complex physical tasks, going far beyond just driving.