Breaking

DrivingBench Puts Four AI Models Behind a Toyota Corolla; GPT-6 Astra Is the Only Finisher

October 7, 2026 at 15:14 EDT

Foundation Models
AI Agents
Robotics

DrivingBench, a new real-world benchmark from Aditya Ramabadran, Simon Mahns and Tobias Gessler, connected four general-purpose vision-language models to a 2022 Toyota Corolla and asked them to drive a cone course at low speed. GPT-6 Astra was the only model-harness system to finish, completing the course on its second attempt in 5 minutes and 22 seconds.

Sept. 30, 2026 - DrivingBench

A Real Car Test for General Models

Four vision-language models controlled a Toyota Corolla on a cone course. GPT-6 Astra alone reached the finish, while latency and perception failures ended the other runs.

4models tested

1model finished

127 mcourse centerline

5:22finish time

Best Course Progress

Astra reached the blue finish zone; no other system reached half the course.

100%

GPT-6 Astra

45%

Claude Fable

11%

Grok 4.6

GPT-5.6 Sol

Closed-Loop Control Path

ObserveCamera frames and telemetry

CommandSteering, speed, duration

MoveCar keeps moving while the model thinks

Replace or StopNew command replaces the old one

What Impressed

A general model controlled a physical car in closed loop.
Astra improved from 49% to a full finish in the same chat.

What Constrained It

Cloud latency made stale camera views a real control problem.
A safety driver, low speeds, and a private course were required.

Continue reading

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