OpenAI’s GPT-6 Astra Just Drove a Real Car Through an Obstacle Course

The driving demonstration highlights the evolving capabilities of general-purpose large language models, not autonomous vehicles. GPT-6 Astra, designed for text and code, drove a real car without specific driving training. This contrasts sharply with specialized self-driving systems from companies like Waymo or Tesla. The new DrivingBench benchmark shows a clear performance gap, with Astra completing the course while older models like GPT-5.6 Sol barely moved.

For Asia, this points to a continuing shift in AI development. Asian companies heavily investing in foundational models, such as those in China and South Korea, must now consider these unexpected multimodal applications. The ability of a general-purpose model to perform physical tasks, even simple ones, expands the competitive landscape. It suggests that a model's utility is no longer confined to its initial training domain.

The thing to watch is how quickly these general-purpose models can move from controlled environments to more complex, real-world interactions. The current test was slow, capped at 3.5 meters per second, and involved a limited field of view. The next test for these models will be adapting to dynamic, unpredictable environments beyond a simple cone course. This will determine their true commercial viability in robotics or physical automation.