AI Model Struggles in Real-World Driving Test

OpenAI's GPT-6 Astra barely completes a parking lot course while other models fail.

By ·

Sep 27, 2026

5 cards · 50s · tap to begin

From · · · 50s

AI Model Struggles in Real-World Driving Test

OpenAI's GPT-6 Astra barely completes a parking lot course while other models fail.

In brief

OpenAI's GPT-6 Astra barely completes a parking lot course while other models fail. Frontier AI models need significant safety improvements before real-world driving. Originally reported by Futurism.

First Test Attempt

A screenshot of researchers using frontier AI models to drive a car.

Three frontier AI models loaded into a Toyota Corolla tried to complete a parking lot course with no human input.

Only One Model Succeeds

GPT-6 Astra completed the course in five minutes after two attempts, covering less than 500 feet.

Most Models Fail Early

All other models failed to pass the first corner, struggling with lane perception and obstacles.

“The car is wider than the camera makes it look”

Safety Concerns Persist

Safety Concerns Persist

Some models refused to drive due to safety concerns, even in empty lots with low speed limits.

“Some models would refuse to drive the physical car sometimes citing safety reasons”

Costly Computing Needed

The test cost $7.74 for 6.6 million tokens, equivalent to 500 times the fuel cost for a 25 mpg car.

The short version

Frontier AI models need significant safety improvements before real-world driving.

Read the original on Futurism

React

Sign in to react and comment.

Comments (0)

Life is short. Keep it sweet. Respect others' opinions and be kind!

    Recommended next

    More decks on openai and related topics.