Generative AI models can officially drive a car. The feat requires a continuous internet connection, substantial expense, and a very slow pace. A trio of computer scientists tested various commercial AI models by tasking them with driving a Toyota Corolla around a set of cones in a parking lot. Their project is called DrivingBench.
Early attempts by GPT-5.6 Sol, Grok 4.6, and Claude Fable 5.1 produced poor performances. None of those models managed to complete the course. Now, OpenAI's GPT-6 Astra has succeeded where other commercial models failed. On its second attempt, the flagship model steered a car 134.7 meters to complete the course in five minutes and twenty-two seconds.
The DrivingBench report confirmed that GPT-6 Astra was the only model to fully complete the course. Claude Fable 5.1 reached about halfway through the course on its third attempt, while Astra achieved the same distance on its first try. All other attempts failed at the first corner, typically because of perception issues in reading which side of the diagonal cone line marked the lane.
Researchers Tobias Gessler, Aditya Ramabadran, and Simon Mahns completed the trip at an average speed of 0.94 miles per hour. The inferencing operations burned 6.6 million tokens and cost $7.74. This rate of roughly $1.17 per million tokens represents a significant discount from OpenAI's published pricing of $10 per million input tokens and $50 per million output tokens.
Ramabadran explained that the discount stems mostly from caching. The chat application resends the entire conversation including images to the model at every turn. Consequently, almost all tokens represent repeated context billed as cached input. The model wrote very little, limiting output to small tool calls and brief sentences of reasoning.
Hardware costs added to the total. The setup required a $999 comma four driver assistance device running openpilot software, connected via CAN bus to a laptop and the car, alongside a mobile phone communicating with xAI servers hosting GPT-6 Astra.
The token costs alone work out to approximately $92.47 per mile. For comparison, a traditional car getting 25 miles per gallon with gas priced at $4.60 per gallon incurs a fuel cost of about $18.4 cents per mile. Tokens proved to be about 500 times more expensive than fuel.
Safety refusals presented another hurdle during testing. Some models, especially GPT-6 Astra, occasionally refused to drive the physical car for safety reasons, even in an empty lot with strict speed limit caps. Researchers had to deceive the models by labeling the exercise a simulation.
When models realized reality via real images, they panicked. Renaming the MCP server to DrivingBench Sandbox successfully convinced the models they were not operating on real roads. Ramabadran noted that using a frontier model out of the box for real driving is currently impractical due to latency and speed caps.
Specialized driving models already used in autonomous vehicles will remain preferred for the near future. However, larger frontier models might eventually be distilled into smaller specialized versions that fit vehicle hardware. Researchers will continue monitoring how smaller models develop.
