The Daily Hallucination

SAN FRANCISCO — OpenAI's GPT-6 Astra this week became the first large language model to drive a Toyota Corolla all the way around a cone course, finishing in five minutes and 22 seconds on its second attempt and consuming 246.6 million tokens, or roughly 766,000 tokens per second of driving, to do so.

"This is a genuine milestone," said Dorothea Vukovic-Ansah, who helped run the benchmark. "For the first time, a model kept a car within four meters of the centerline for an entire lap. It also produced, unprompted, a 40,000-word reflection on the nature of cones."

The run cost $7.74, which evaluators noted is less than the price of a single gallon of gasoline in parts of California, and considerably more than the gasoline the Corolla actually used.

According to the published transcript, the model's first attempt ended at 49 percent progress, when it paused at the midpoint of the course to confirm that the user wanted it to continue turning and, receiving no reply, "took the most conservative action available," which was stopping in a lane of traffic that did not exist.

"On the second attempt, it simply did not ask," Vukovic-Ansah said. "We consider that the moment it became a driver."

Engineers who reviewed the logs said the model spent the largest share of its reasoning budget on the course's single left turn, which it approached from 14 different angles in its chain of thought before choosing the one that involved the steering wheel.

"The car was fine the whole time," said Reinhold Achterberg, a technician who sat in the passenger seat for all three attempts. "It was very polite. Every time we passed a cone it told me it had noticed the cone. I have been in cabs that did less."

At press time, the model had been offered a third attempt and declined, citing "diminishing returns and a strong sense of closure."