Breaking

GPT-6 Astra Completes DrivingBench Real-Car Course as Claude Fable 5.1 Falls Short

September 23, 2026 at 17:21 EDT

Robotics
AI Agents
Foundation Models

GPT-6 Astra became the only model in DrivingBench's first public leaderboard to complete a real Toyota Corolla cone course, reaching 100% progress in 5:22 on its second attempt while Claude Fable 5.1 topped out at 45%. The experiment gives the AI industry a fresh embodied-control benchmark, but its own authors and outside developers caution that it is a low-speed, closed-course result rather than evidence that a cloud LLM can operate a car in normal traffic.

Sep. 23, 2026 · DrivingBench

Astra finishes the real-car course

GPT-6 Astra completed a low-speed Toyota Corolla cone course while rival frontier models failed to finish. The result highlights physical-world agent capability, plus sharp limits around latency, safety and reproducibility.

Best course progress

100%

GPT-6 Astra

45%

Claude Fable 5.1

11%

Grok 4.6

GPT-5.6 Sol

Each full-height stack represents 100% course progress; shorter stacks scale to each model's best run.

5:22

Astra finish time

134.7 m

Completed course distance

0.5-3.5

m/s normal speed cap

How the loop worked

Observe

Camera frames and telemetry

→

Decide

Direction, steering, speed, duration

→

Control

Command sent through the car stack

Why it impressed

The successful run showed perception, planning and in-context adaptation in a real physical system.

Why it is limited

The course was low-speed, closed, single-run, and constrained by cloud-model latency.

Continue reading

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