GPT-6 Astra Drives a Real Corolla Through DrivingBench

A GPT-6 model called Astra just drove a real Toyota Corolla through an entire cone course, and on its second attempt it finished the full route in 5 minutes 22 seconds. Its first attempt only got 49% of the way.

That run put Astra ahead of every other frontier model on the same benchmark, which is why the Polymarket odds on OpenAI declaring AGI moved.

What Happened

In the DrivingBench test, Astra's first attempt ended at 49% of the course. The second attempt finished 100% in 5 minutes 22 seconds.

It was the only frontier model to cross the line. Claude Fable 5.1 stalled at the halfway point on its third attempt. Grok 4.6 and GPT-5.6 Sol never reached the finish either.

The task itself was narrow. The car made one pass through a cone course, taking one command per action, with a human supervisor sitting in the seat ready to hit the brake.

Why It Matters

What Astra demonstrated is real-time sequence planning, not the broad, general capability that AGI is usually defined as. The benchmark is a cone course, driven once, with a person able to stop the car at any moment.

The betting market treated it that way. Polymarket still lists the event "OpenAI announces it has achieved AGI before 2027?" at 19%, which is only a 4-point rise after the news broke. The traders calling Astra impressive and OpenAI's AGI claim not yet earned are, for now, pricing it the same way.