OpenAI's GPT-6 Astra, released in early September, was the only one of four leading general-purpose large models to complete a real car-driving test, according to Chinese automotive outlet D1EV. In an empty parking lot in California's Bay Area, a 2022 Toyota Corolla crept through a cone-marked course at slower than walking pace, while a human safety driver kept their hands off the wheel and hovered one foot over the brake. The Corolla needed 5 minutes 22 seconds to finish the 134.7-metre course and stop inside a box of blue cones.
The test was run by DrivingBench and published on 21 September. The other three participating models initially refused to drive a real car, citing safety, even after researchers explained that the lot was empty, the speed was minimal and a human was ready to brake. The breakthrough came only when the tool server linked to the vehicle was renamed “DrivingBench Sandbox”, so the models treated the task as a simulation.
On the car, the models were genuinely blocked by rows of angled cones. They struggled to determine which side of the road they were on, and while they thought, the car continued on its previous command. Each failure was followed by a written reflection and another attempt, up to three times. GPT-5.6 Sol stopped after 6% of the course, Grok 4.6, which reflected that “the car is wider than it looks on camera”, reached 11%, and Claude Fable 5.1 thought so long that it covered only 45%.
GPT-6 Astra also managed only half the course on its first attempt, admitting in its reflection that it had “prematurely judged the vehicle to be straight”. It then slowed around corners and shortened its instructions, eventually completing the 134.7 metres in 5:22 while using 6.6 million tokens, which cost about US$7.74 at official prices.
Who actually drove the car
For all the AI involved, GPT-6 Astra did not turn the steering wheel. The Corolla ran openpilot, comma.ai's open-source assisted-driving software, which normally handles lane centering and adaptive cruise control at Level 2. The large model supplied the direction and speed decisions, and a control program translated those into steering angle and acceleration requests, which openpilot sent to the car over the CAN bus.
A human pressed the RES button on the steering wheel before each run and kept a foot ready over the brake. Openpilot was in contact with the vehicle 100 times a second, while Astra looked at the road only once every five or six seconds; in that gap, openpilot and the car had already exchanged 500 to 600 messages.
Why the test matters for China's ADAS developers
D1EV stresses that this is not a sign that general-purpose AI is ready for daily driving. At 60 km/h (about 37 mph), a car covers 83 metres in five seconds, and an emergency stop or a pedestrian stepping out will not wait for a large model to think.
The bigger effect, the article argues, is on how ADAS models are trained. Waymo already uses Gemini to help train models that understand driving scenes, then distills the capability into small on-car models. Nvidia's Alpamayo 1 is a vision-language-action model with reasoning ability, like a driving instructor that watches road video and proposes the next move. Alibaba's Qwen-Drive-1.0 takes a different approach, attaching a 3D perception module and a planning module to a pretrained Qwen3.5-4B model, though D1EV describes it as still laboratory-level.
The 134.7-metre run therefore does not mean the ADAS (advanced driver-assistance systems) industry is about to be replaced. Developers who spent heavily to teach models road understanding can now buy part of that capability more cheaply from general-purpose models, says D1EV, and focus their own work on adapting the model to the vehicle and proving it behaves safely. The practical skill of engineering remains the real job.
