GPT-6 Astra completes real-car driving test as Grok and Claude fail

By Splitfeed EditorialLast updated Oct 7, 2026, 20:30 UTC

Between October 5 and 7, 2026, three engineers conducted an experiment in which they placed three general-purpose AI models, GPT-6 Astra, Claude, and Grok, in control of a real Toyota Corolla and asked each to navigate a cone course. GPT-6 Astra was the only model to complete the task. Grok failed outright, and Claude also did not finish successfully. The Wired report adds a notable detail about the destination that framed the exercise, suggesting the drive had a practical, if informal, goal. No injuries, vehicle damage, or specific timing and accuracy metrics appeared in the available reporting.

The significance of the test lies in its setting: a physical road environment rather than a simulator, which raises the stakes considerably for evaluating what a general-purpose language model can and cannot do when connected to real hardware. General-purpose models were not designed primarily for vehicle control, so any successful navigation of a physical course represents a meaningful capability marker. At the same time, outlets were quick to note that GPT-6 Astra's performance still falls short of dedicated autonomous driving systems like Waymo, which have accumulated millions of real-world miles under specialized development programs. The genuine dispute here is whether a result like this signals that general-purpose AI is converging on specialized systems, or whether the gap between completing a cone course and operating safely in traffic remains so large as to make the comparison premature.

None of the three camps, Pro-AI, Anti-AI, or Middle Ground, had published reactions to this story at the time of writing. Pro-AI voices would typically treat GPT-6 Astra's success as evidence that general-purpose models are developing physical-world competence faster than skeptics expected, and would likely point to the failure of competitors as a sign that OpenAI is pulling ahead. Anti-AI voices would typically focus on what the test does not prove, arguing that a controlled cone course is far removed from the unpredictability of public roads, and that framing this as a driving capability risks encouraging premature deployment. Middle Ground observers would likely acknowledge the result as a genuine incremental step while insisting that the Waymo comparison is the more relevant frame, and that purpose-built systems with extensive safety validation still set the appropriate standard.

The clearest thing to watch next is whether any of the three teams, or independent researchers, attempt a more rigorous follow-up with documented metrics such as completion time, cone strikes, and speed, or whether the experiment is replicated in open traffic conditions rather than a controlled course. A formal technical writeup from the engineers involved would also clarify how the models were interfaced with the vehicle, which is currently absent from the reporting.