DrivingBench Reworked AI Prompts Before One Model Finished a Corolla Course

The only AI model to finish DrivingBench’s parking-lot driving course first resisted taking the wheel. In a new interview with 404 Media, team member Aditya Ramabadran said the researchers spent hours changing their prompts before GPT-6 Astra would consistently drive a Toyota Corolla. How they got that result matters as much as the finish.

Before the first turn

DrivingBench connected four models—GPT-6 Astra, Claude Fable 5.1, Grok 4.6 and GPT-5.6 Sol—to a Corolla’s steering, accelerator and brakes. The assignment was to stay between cones on a low-speed parking-lot course and finish in a marked parking area. The team wanted to see whether off-the-shelf AI models could control a real car, not build a system for everyday driving.

Getting the models to accept that assignment became part of the experiment. Ramabadran told 404 Media that most initially refused when they recognized they were being asked to drive a car. Calling it a simulation sometimes worked, he said, but the models could see images of the real parking lot.

We ended up having to call everything a sandbox. And with that prompt, it's able to consistently drive the car and like never refuse to do that.

Aditya Ramabadran, DrivingBench member, speaking to 404 Media

A model at the controls, people in the car

A modified Comma system linked the Corolla to a laptop running the models. Two cameras aimed at the road supplied visual input; the models issued driving commands through the connection. That setup gave a chatbot a way to act on a physical vehicle, rather than merely describe what it saw.

It was not an unattended drive. A team member prompted the model from inside the car. A person pressed a steering-wheel button to start movement, monitored the vehicle and kept a foot over the brake. Those precautions distinguish the test from letting a model take a passenger onto the road.

One finish, several early failures

GPT-6 Astra was the only tested model to complete the course. It finished on its second attempt, in five minutes, according to Futurism’s account of the team’s results. Grok 4.6, Claude Fable 5.1 and GPT-5.6 Sol went only a few meters. The researchers said many failed attempts stalled at the first corner, where models struggled to read which side of a diagonal row of cones formed the lane.

The successful run covered less than 500 feet at about 0.94 mph. Finishing that short, slow course shows something narrower than coping with ordinary traffic.

The pause between seeing and steering

Ramabadran identified response time as a reason to test models in a moving car. His example was simple: if a model takes 10 seconds to respond while a vehicle travels about one meter per second, the vehicle moves roughly 10 meters before the next answer. Even a low-speed course can expose a problem that a static image test would miss.

The team has posted code, prompts and videos for others to inspect. The interview describes how the wording changed; the posted materials let readers examine the driving instructions and footage. Neither makes this supervised run proof that a general-purpose chatbot can drive safely without a person ready to intervene.