OpenAI’s GPT-6 Astra Just Drove a Real Car Through an Obstacle Course
Trending Topics Jakob Steinschaden ● Covered by 3 sources
OpenAI’s GPT-6 Astra drove a Toyota Corolla through a cone course. It beat Anthropic and xAI, but only in a controlled test with a brake-ready human onboard.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI’s GPT-6 Astra has already written emails and shipped code. Now it can also guide a Toyota Corolla through a cone course, at least in a benchmark built for that exact stunt.
The test, called DrivingBench, put Astra in front of other big-name models. According to the project’s results, it was the first major language model to steer a real car all the way through the course, while systems from Anthropic and xAI fell short.
The setup was intentionally simple and a little strange. Developers Aditya Ramabadran, Simon Mahns and Tobias Gessler fitted a Toyota Corolla with comma.ai hardware that let software control the steering, accelerator and brakes. The models didn’t take direct control of the car. They drove one command at a time inside an ongoing chat session, running in their usual coding environments — GPT-6 Astra in Codex, Claude Fable 5.1 in Claude Code, and so on.
The route itself was a roughly 130-meter loop of traffic cones in a parking lot. Each model had up to three tries, and the scoring measured how far it stayed on the centerline without drifting more than four meters away. Safety limits were built in too: speed was capped at 3.5 meters per second, and a person in the car could hit the emergency brake at any moment.
Astra finished the course on its second attempt in 5 minutes and 22 seconds. That was enough for 100 percent completion. Claude Fable 5.1 reached 45 percent, Grok 4.6 got to 11 percent, and GPT-5.6 Sol managed just 6 percent across all attempts. The split between the two OpenAI models is the sharpest part of the result. One barely left the starting line. The other went the distance.
The team published videos, logs and code on GitHub, but the benchmark comes with some obvious limits. Each model was evaluated only once, the attempts all happened in the same conversation, the camera view was restricted and the steering was tuned to be cautious. The researchers also stress that the project is just research software, with no ties to comma.ai, Toyota or the AI companies involved.
My take — AI-written commentary, not fact-checked reporting
This is a nice demo, not a self-driving breakthrough, and the internet should stop pretending those are the same thing. The real story is simpler: general-purpose models are getting better at acting in the world, which is exactly why they need tighter guardrails, not more hype. Also, a car creeping through cones at walking pace is not a robotaxi, no matter how shiny the benchmark sounds.
Read more about this at: Trending Topics