GPT-6 Astra was the only model to complete the DrivingBench course in a Toyota Corolla
GPT-6 Astra completed a 134.7-metre cone course in 5 minutes 22 seconds. The test took place in an empty car park with a human operator ready to brake.
In the DrivingBench experiment, GPT-6 Astra became the only one of four tested general-purpose language models to complete a driving course. The test used a Toyota Corolla on a cone-marked route in an empty car park.
The setup was not a production self-driving system. A laptop sent camera frames and telemetry to the model, which returned steering, acceleration and braking commands through a comma four device and openpilot. A human operator remained in the vehicle throughout the runs with a foot ready over the brake.
On its second attempt, GPT-6 Astra covered 134.7 metres in 5 minutes 22 seconds and issued 24 motion commands. The run used 6.6 million tokens and was priced at $7.74. Claude Fable 5.1 reached 45% of the course at best, while Grok 4.6 reached 11% and GPT-5.6 Sol reached 6%.
The project’s authors linked most failures to spatial perception: models misread which side of a diagonal line of cones was the driving lane. The evaluation also has clear limits. Each model received up to three attempts in one continuous conversation, and the course was driven slowly in a controlled setting. The result therefore does not show that language models are ready to drive on public roads.
Добавьте Vzglyad.uz в предпочтительные источники Google. Это бесплатно и занимает 5 секунд, а для нас — лучшая поддержка и мотивация работать ещё лучше.