AI Driving Test: How Language Models Drive Cars

Enthusiasts conducted an experiment by placing popular language models behind the wheel of a car. The goal was to debunk the myth that AI is incapable of controlling a vehicle.

A Toyota Corolla was selected for testing. The neural networks were given control over the accelerator and brakes, with a maximum speed limit on turns set at 13 km/h.

The test results showed varying degrees of success among the models. Route success was evaluated as a percentage of the distance covered or successful attempts.

Language model testing results

• ChatGPT-6 Astra successfully completed the route on the second attempt, making a mistake on a sharp turn during the first try
• Claude Fable 5.1 achieved a result level of 45%
• Grok 4.6 managed to cover only 11% of the distance
• GPT-5.6 Sol could only drive 6% of the route