GPT-6 Astra has gained the ability to drive a car
Points and comments are a snapshot, not live.
A benchmark called DrivingBench gives frontier LLMs control of a real car to navigate a cone course.
DrivingBench evaluates frontier AI models by giving them control of a Toyota Corolla's steering, accelerator, and brakes on a fixed cone course. Models are ranked by progress along the course centerline while staying within 4 meters, distance traveled, finish time, commands issued, and token cost. Runs are recorded with video traces. The benchmark tests out-of-the-box vision and reasoning in a real driving task, not specialized driving models.
What commenters are saying
Commenters found the benchmark cool and novel. Some noted Astra initially refused to drive until the MCP was renamed to "DrivingBench Sandbox," highlighting how models are more compliant when told they are in a sandbox. Several discussed latency issues, noting the models drive slowly (5 minutes vs 15 seconds for a human). Others debated whether general LLMs will replace specialized self-driving stacks, with some pointing out that bitter lesson suggests general architectures may eventually win, while others argued specialized systems still outperform today.