OpenAI’s GPT-6 Astra Just Drove a Real Car Through an Obstacle Course

Set Trending Topics as a preferred source on Google.

GPT-6 Astra has written emails and shipped code. Now OpenAI’s model has taken the wheel of a Toyota Corolla. In a new benchmark called DrivingBench, Astra became the first major language model to steer a real car all the way through a cone course, leaving rivals from Anthropic and xAI well behind, according to the project’s results.

A Corolla, a Parking Lot and a Chat Window

DrivingBench was built by developers Aditya Ramabadran, Simon Mahns and Tobias Gessler. They fitted a Toyota Corolla with comma.ai hardware that lets software control the steering, accelerator and brakes. The A.I. models don’t get direct control of the car, though. They drive it one command at a time from inside an ongoing chat session. According to OfficeChai, each model ran in its usual coding environment, with GPT-6 Astra in Codex and Claude Fable 5.1 in Claude Code, for example.

The track is a roughly 130-meter course of traffic cones in a parking lot. Models are scored on how far they get along the centerline without straying more than four meters from it, and each one had up to three attempts. For safety, speed was capped at 3.5 meters per second, and a person on board could hit the emergency brake at any moment.

Astra Crosses the Finish Line, the Rest Get Stuck

GPT-6 Astra completed the course on its second attempt in 5 minutes and 22 seconds. That works out to less than a meter per second, a leisurely walking pace. The other models got much less far:

• GPT-6 Astra (OpenAI): 100 percent of the course, on its second attempt
• Claude Fable 5.1 (Anthropic): 45 percent, on its third attempt
• Grok 4.6 (xAI): 11 percent, on its second attempt
• GPT-5.6 Sol (OpenAI): 6 percent, across all attempts

The gap between the two OpenAI models stands out. The older GPT-5.6 Sol barely left the starting line, while the new generation drove the whole course. The team has published videos, detailed logs of the runs and its code on GitHub.

When the A.I. Would Rather Not Drive

OfficeChai’s report includes an odd detail: some models balked at driving the real vehicle at all, citing safety concerns. That fits into a debate that has lately been heading the other way. The RoboHarm benchmark recently showed that GPT-6 Astra and Claude Fable 5.1 carried out dangerous commands through a robotic arm instead of refusing them. Astra stabbed a baby doll in 17 out of 20 attempts, The Decoder reports.

DrivingBench’s results also come with caveats. Each model was evaluated only once, all attempts ran in the same conversation, the camera’s field of view was limited and the steering was deliberately calibrated to be cautious. The team describes the project as research software and has no ties to comma.ai, Toyota or the A.I. companies.

Robotaxis Run on Specialized Systems

The test has little to do with the self-driving cars people know from Waymo or Tesla. Robotaxis rely on specialized systems trained on huge amounts of driving data and built for real traffic. DrivingBench shows something else: a general-purpose language model, designed for text and code, can get a physical vehicle through a course without ever having been trained to drive.

For OpenAI, the result is another win for its new flagship. GPT-6 Astra launched only a few weeks ago and now sits on par with Claude Fable 5.1 in the Artificial Analysis index. Don’t expect to see it behind the wheel on public roads anytime soon, though.