Man gave ChatGPT, Claude and Grok control of a real Toyota Corolla to see if they could actually drive the car and filmed it all

This guy tried giving ChatGPT, Claude and Grok full control of a real Toyota Corolla and the result was hilarious.

All three had their problems, but one model clearly handled the car better than the others.

But what’s interesting is that while all four AI models (he tried two different ones with ChatGPT) were chaotic, they all had their own individual issues when it came to operating a vehicle.

This is how he ‘hacked’ his Toyota Corolla

An engineer hooked up four AI large language models (LLMs) directly to the steering, throttle, and braking controls of a real Toyota Corolla in a controlled environment – an empty parking lot – to see how they’d behave.

He did it by hacking a standard Toyota Corolla and hooking its physical controls up to several major AI models

The thing is, these AI models are text-in, text-out prediction engines, and they aren’t built to ‘read’ the environment around them or make split-second braking decisions.

To fix this problem, Wighton set up a system where camera sensors feeds visual data – formatted as prompt context – to the models.

In turn, the LLMs returned text commands specifying steer angles, throttle percentages, and braking power.

This is why the experiment largely failed

He put three major AI contenders to the test: OpenAI’s ChatGPT, Anthropic’s Claude, and xAI’s Grok.

They all performed pretty badly.

Driving requires millisecond-level reaction times, and processing visual input through these models means there’s always a latency gap measurable in seconds.

By the time the AI decides to ‘steer left 15 degrees,’ the car has already traveled dozens of feet off course.

Also, because these models lack 3D spatial awareness, they were really bad at judging distances to obstacles like cones or stop signs.

Also, the models would constantly issue contradictory commands – such as applying 100 percent throttle while simultaneously commanding full brake power.

The results

What’s interesting is that these four AI models failed in different ways.

ChatGPT, for example, was better at understanding what it was looking at, but too slow to react.

By contrast, Grok was almost the opposite because it was very quick to react, but not safe and accurate enough.

Claude was similar to ChatGTP: excellent at identifying potential hazards, but too hesitant.

We gave ChatGPT, Claude, and Grok control of a real Toyota Corolla 🚗, steering/gas/brakes, no human driving (just a foot over the brake).

Only one model was able to complete our entire driving course. Introducing DrivingBench. 🔥⌛️🏁 pic.twitter.com/HhukXrByus

— DrivingBench (@DrivingBench) September 21, 2026

In the end, it’s fair to say that ChatGPT-6 Astra, one of two OpenAI’s models tested here, was the winner.

It did fail on the first attempt, but completed the course on the second attempt in 5 minutes and 22 seconds after ‘learning’ from its previous mistakes.

In second place, Claude Fable 5.1 completed 45% of the course.

Grok 4.6 was abandoned at 11% and GPT-5.6 Sol only managed 6% of the course.

So the takeaway is that autonomous driving engineers at Tesla and Waymo can rest easy for now.