Breaking

GPT-6 Astra Completes DrivingBench Real-Car Obstacle Course as Rival Models Fall Short

September 23, 2026 at 11:36 EDT

AI Agents
Robotics
Foundation Models

GPT-6 Astra completed a real-world DrivingBench cone course in a Toyota Corolla, becoming the only model on the benchmark's published leaderboard to finish the run. The test gave frontier language models control of steering, acceleration and braking through a constrained tool interface, with a human operator supervising from the driver's seat.

September 2026 · DrivingBench

GPT-6 Astra finishes a real-car AI driving test

The model completed a low-speed Toyota Corolla cone course while other tested models stopped short. The result points to stronger tool-use and physical-world adaptation, but only inside a tightly supervised closed-course setup.

100%

Astra best progress

5:22

Finish time

Motion commands

134.7 m

Completed distance

Leaderboard progress

100%

GPT-6 Astra

45%

Claude Fable 5.1

11%

Grok 4.6

GPT-5.6 Sol

Astra was the only tested model to reach the finish zone.

What worked

Astra adapted after an earlier partial run, kept speeds at or below 0.8 m/s, and refreshed commands often enough to navigate the course.

What limits it

The run used one closed course, one Corolla setup, low speeds, and a human supervisor ready to brake at every moment.

Control loop

Observe frames→Set motion→Car moves→Adapt again

Commands replaced the active motion, so timing and repeated observation were part of the test.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.