Engineers Let GPT, Claude and Grok Drive a Real Toyota Corolla; Only One Finished the Trip

Three engineers wired GPT, Claude and Grok into the controls of a real Toyota Corolla and let each model attempt an actual drive, with only one of the three making it to the In-N-Out destination. The experiment, reported by Wired, turned a consumer sedan into a testbed for whether general-purpose language models can handle the messy, split-second decisions of public roads.

The setup matters more than the stunt framing suggests. Rather than training a purpose-built driving policy, the team handed the models control of a physical vehicle and let their reasoning run against live traffic, lane markings and intersections. Each model got the same car and the same route, so the differences in outcome trace back to how the systems interpreted what they saw and when they chose to act.

That only one model completed the run is the headline result, but the failure modes of the other two are the more useful data. Language models are built to produce plausible next tokens, not to respect stopping distances or the reaction time a pedestrian assumes. A model that hesitates for a second while it reasons through an ambiguous scene is behaving exactly as designed, and exactly wrong for a car moving at speed.

The experiment sits alongside a broader push to use foundation models as the reasoning layer in robotics and autonomy. Waymo and Tesla rely on specialized stacks tuned for driving, while this test asks whether off-the-shelf models can substitute. The answer so far looks like a qualified no: capable of parsing a scene and proposing an action, but not yet reliable enough to be trusted alone with a two-ton vehicle.

The engineers have not published a full safety or reliability breakdown, which leaves the key question open. How much of the Corolla's run was the model driving, and how much was human supervision, remote override or a safety driver ready to grab the wheel? Without that split, the result is a compelling demo rather than evidence about model competence.

What to watch next is whether the team releases per-model failure logs and repeats the route under controlled conditions. If one model consistently outperforms the others on the same hardware, that says something concrete about instruction-following and spatial reasoning under uncertainty, the same skills that matter for warehouse robots, drones and any agent expected to act in the physical world. Until those numbers exist, treat the drive as a signal about where language models are strong and where they still are not.