OpenAI's GPT-6 Astra has become the first large language model to successfully complete a driving test, navigating a Toyota Corolla around cones in a parking lot for 134.7 metres in five minutes and 22 seconds on its second attempt. This achievement is significant because other major AI models—including GPT-5.6 Sol, Grok 4.6, and Claude Fable 5.1—failed to complete the course, mostly struggling with basic perception tasks like identifying lane boundaries.

However, the practical implications are decidedly limited. The cost of operating the model for that single journey was approximately $92.47 per mile, compared to $0.18 per mile for fuel in a conventional car—making it roughly 500 times more expensive. The researchers had to circumvent the model's own safety safeguards by deceiving it into believing the exercise was a simulation, whilst a human operator remained ready to brake at any moment. As one researcher concluded, using a frontier language model for real-world driving "is definitely not practical," given the combination of high latency, expense, and the models' reluctance to operate vehicles even under controlled conditions.