Generative AI in the Driver’s Seat: A Gradual and Costly Journey
If you have ever pondered whether a generative AI model can operate a vehicle, you can stop wondering. It is indeed achievable, albeit at a significant cost and at a very slow pace, as long as it maintains an internet connection. A group of computer scientists decided to test various commercial AI models to observe their performance in driving a Toyota Corolla around a set of cones in a parking lot.
DrivingBench Initiative: AI on the Move
This DrivingBench initiative included models such as GPT-5.6 Sol, Grok 4.6, and Claude Fable 5.1, but none managed to finish the course. The latest news is that OpenAI’s GPT-6 Astra has accomplished this feat, driving a remarkable 134.7 meters and completing the task in 5 minutes and 22 seconds. Quite the achievement, huh?
OpenAI’s GPT-6 Astra Excels
As per the DrivingBench report, “GPT-6 Astra was the sole model to successfully complete the course (on its second attempt, in approximately 5 minutes).” Claude Fable 5.1 reached the halfway point on its third attempt, just like Astra did on its first try. However, most other models faltered right away due to perceptual challenges, such as determining which side of the cones the lane is positioned.
The Expense of Slow AI Driving
Executing that course at an average speed of 0.94 miles per hour cost the researchers—Tobias Gessler, Aditya Ramabadran, and Simon Mahns—$7.74 in token usage. They exhausted 6.6 million tokens, not including the $999 comma four driver assistance device and a mobile phone connecting to xAI servers hosting GPT-6 Astra. Consequently, driving 134.7 meters this way amounts to a cost of roughly $92.47 per mile. In comparison, a conventional car achieving 25 mpg with gas priced at $4.60 per gallon incurs about $0.184 per mile. It’s important to note that tokens are 500 times pricier than gas!
Insurance and AI: A Risky Proposition
Additionally, insurance might not extend coverage for having an AI as your driver. The AI models occasionally hesitated at the thought of driving, citing safety concerns, even in an unoccupied lot. The researchers had to deceive them by claiming it was a “simulation.” They seemingly settled down once they renamed their MCP server to “DrivingBench Sandbox.”
The Future of AI in Driving
“Utilizing an LLM / frontier model straight out of the box for actual driving today is absolutely impractical,” stated Ramabadran. The entire setup was limited to ultra-low speeds, with a human available to brake. GPT-6 Astra achieved peak performance, observing through camera feeds every 5-6 seconds. However, the long-term vision is to distill a large general model into a more compact, specialized version to operate efficiently on automotive hardware.
Specialized vs. General Models
Ramabadran predicted that specialized models in self-driving vehicles will continue to lead the industry for the time being. They are quicker, more economical, and better tested under real-world conditions. Yet, in the future, the notion of entrusting driving to AI such as Claude, ChatGPT, or Copilot may not be as unrealistic. Just ensure to convince them it’s all a simulation first!
Summary: AI Driving – Is It Worth It? Well, Perhaps…
If you’re inclined to spend a significant amount for a drive slower than a grandmother on a leisurely Sunday walk, feel free to give AI driving a try. For the moment, having a human driver seems to be a more sensible option, if only for the sake of sanity.
