---
title: "Uh Oh, ChatGPT Can Drive a Car (And Grok Crashed Immediately)"
date: "2026-10-05"
modified: "2026-10-05"
authors:
  - name: "Adam Ismail"
    job_title: "Senior Editor"
    link: "https://www.thedrive.com/authors/adam-ismail.md"
url: "https://www.thedrive.com/news/uh-oh-chatgpt-can-drive-a-car-and-grok-crashed-immediately"
categories:
  - "Car Tech"
  - "News"
  - "Self-Driving Tech"
---

# Uh Oh, ChatGPT Can Drive a Car (And Grok Crashed Immediately)

![DrivingBench footage of car driving on course with AI chat overlaid](<https://www.thedrive.com/wp-content/uploads/2026/10/Screenshot-2026-10-05-124453.jpg>)
*\<em\>DrivingBench, composite by The Drive\</em\>*

What happens when you take a "frontier" (read: general-purpose) AI model like Claude or GPT and ask it to get behind the wheel and drive a car? Sure, some self-driving systems utilize artificial intelligence in one sense or another, but they're specifically built for the task. Researchers wanted to see how well your typical AI agent handles the task of driving, via [Comma vision hardware](<https://www.thedrive.com/news/a-1000-hands-free-driving-gadget-drove-an-old-prius-coast-to-coast-99-autonomously.md>) and OpenPilot. The results surely won't leave anyone racing to replace themselves in the driver's seat anytime soon, but they are eye-opening.

Aditya Ramabadran, Simon Mahns, and Tobias Gessler created [DrivingBench](<https://drivingbench.com/>): A simple test for Claude, GPT, and Grok models to navigate a short point-to-point course around a parking lot with a few curves, outlined by small cones. Each agent was issued the same command, with the same guidelines and directions to share what they're seeing and doing after every step, as well as full control of a Toyota Corolla. They regularly stop along the way, because the commands are being sent away to data farms over a wireless hotspot, and all of this happens within one continuous chat. They have three attempts to get it right.

What the DrivingBench crew discovered is that frontier models, frankly, are not good at this. Of the 11 runs conducted across four different agents, eight failed to complete more than 11% of the course, crashing out in the first turn. Each session has a video attached to it, with the chat receipts scrolling in real time so you can see the AI working through it—or, at least, trying to.

![Screenshot of a video of ChatGPT "driving" a car.](<https://www.thedrive.com/wp-content/uploads/2026/10/Screenshot-2026-10-05-124504.png>)
*A screenshot of a DrivingBench trial with a GPT-6 Astra conversation so you can see how the AI is thinking through navigating the course. *DrivingBench**

Grok initially thought a gap between the mini-cones forming the outer boundary was a gate it was meant to pass through; it drove straight off in its first attempt, ending after two commands. One GPT model failed in part because invented a rule that cones on one side of the course were all the same color, which the initial, human-written command explicitly warned was not the case. And many of the agents simply failed to turn enough, because they couldn't work out how much steering angle was required to round the bends.

Reading through these one-sided conversations is fascinating, because as clever as these agents can be at certain tasks, they're completely in over their heads when it comes to driving. All except one, anyway, because GPT-6 Astra was the only agent to complete the course fully, on its second attempt. The route it took was a little wonky as you can see below, heavily hugging the long right-hander before almost driving off the course on the outside just before the finish box, but a win's a win.

![AI driving course layout.](<https://www.thedrive.com/wp-content/uploads/2026/10/Screenshot-2026-10-05-122235.png>)
*This screenshot shows the path of every attempt through the DrivingBench course. It's static here but shows you the AI agents' routes in real time [on the website](<https://drivingbench.com/#course>). *DrivingBench**

Per DrivingBench, Astra's successful run cost $7.74—nearly four times the charge (and way more tokens) than its prior attempt that made it half as far. Some say AI will kill us all; this experiment demonstrates that a very fast way to make that happen is to let the world's "best" agents drive our cars.

*Got a tip? Reach out to tips@thedrive.com*

## Author
Adam is a Senior Editor at The Drive, concentrating on curating and developing the site's slate of daily journalism, from initial pitch to final copy edit. When he's not at his desk researching leads and working with writers, he's testing vehicles or reviewing racing games, new and old. He joined The Drive in 2023, following seven years of covering consumer technology and cars. A car lover and gamer since his earliest memories, Adam came to know the automotive world through video games. First it was Daytona USA, Sega Rally, and Ridge Racer; then came Need For Speed, which inspired an encyclopedic knowledge of '90s supercars, and, of course, Gran Turismo. If not for them, he'd assuredly be doing something else right now. Adam is based in the suburbs outside Philly, where there are ample roads for his hot hatch to stretch its legs. In his car, he's probably listening to indie rock, drum and bass, or Yankees baseball.

### Author social links  
[Bluesky](<https://bsky.app/profile/aismail.bsky.social>)  
[Instagram](<https://www.instagram.com/pioneerspine/>)  
[LinkedIn](<https://www.linkedin.com/in/aismail1/>)