DrivingBench is a research project that evaluates whether frontier language models can control a real car through natural-language-driven tool calls. Its benchmark puts models in control of a Toyota Corolla’s steering, accelerator, and brakes on a fixed course marked with cones, with a human supervisor ready to intervene. The project publishes a leaderboard comparing model performance across up to three attempts conducted within one continuous chat. Results include course progress, distance, finish time, commands, token use, and estimated cost, alongside trajectory traces and video replays that help visitors inspect each run.

DrivingBench also provides information about its evaluation methods, course, and research through its website, report, and linked paper. The project invites questions, feedback, and requests from teams interested in having their models evaluated. It describes itself as research software and cautions users to use it at their own risk. DrivingBench states that it is not affiliated with or endorsed by comma.ai, openpilot, Toyota, or the model and chat-application makers featured in its results.