The large-scale artificial intelligence language model OpenAI GPT-6 Astra is the first language model of its kind to cope with car driving. We’re still a long way from a mature self-driving system: cars move extremely slowly, and API access isn’t cheap, even accounting for caching.
Image source: drivebench.com
As part of the programme, the experiment was carried out by experts bridge. The OpenAI GPT-5.6 Sol, SpaceXAI Grok 4.6 and Anthropic Claude Fable 5.1 models simply couldn’t complete the test track, with only the OpenAI GPT-6 Astra driving a Toyota Corolla along the 134.7m route on its second attempt – crossing the car park between the cones in 5 minutes and 22 seconds. The rest of the models gave up before even the first turn, with most unable to determine which side of the diagonal line the first row of cones was the traffic lane on.
The car traveled around the track at about 1.5 km/h, using 6.6 million tokens and costing $7.74. Therefore, the cost of 1 million tokens is $1.17, which is significantly lower than the official price of OpenAI: 1 million input tokens are $10, and 1 million output tokens are $50. We managed to save money thanks to the caching mechanism. “Because the chat application resends the entire conversation history (including images) to the model at every step, almost all tokens represent repeated context and are charged as cached inputs ($1 per 1 million tokens instead of $10). During the communication process, the model generates minimal text — only sentences for brief reasoning to the tool and a few sentences with answers or responses.”explains the author of the project.
The cost also includes the $999 Comma Four driver assistance system, which installs openpilot software, connects to the laptop and car via CAN bus, and accesses a smartphone running GPT-6 Astra’s SpaceXAI server. Taking into account the cost of accessing the model via API, its operating cost is US$57.46 per 1 km. The average fuel consumption is 9.4 liters per 100 kilometers, the US price is US$1.22 per 1 liter, and the driving cost per kilometer is US$0.114. So the cost of driving a car using OpenAI GPT-6 Astra is more than 500 times higher than the cost of fuel, and you also have to pay when the car is traveling at 1.5 km/h.
The researchers admitted that they actually tricked all the AI models into not refusing to take control of the car for safety reasons—they were told the car was in a simulated state. In some cases, the model realized the situation was real and began to panic after seeing the real image. The most effective approach is to rename the MCP server to DrivingBench Sandbox – this name is a convincing argument to the AI that everything happens in isolation.
GPT-6 Astra, which showed the best results, accessed the camera profile analysis tool every 5-6 seconds. In one attempt, Fable 5.1 directly controlled the car for only 31 of 190 seconds. The rest of the time it simply “thought” and the car stood still. The study’s authors concluded that an effective strategy in the long term would be to train a powerful general-purpose AI model and then refine it into a compact, specialized version that can run on cars and is efficient enough to work in real-world conditions.
If you find an error, select it with your mouse and press CTRL+ENTER.
