DrivingBench

DrivingBench is an open-source research project and benchmark that tests whether frontier, general-purpose large language models (LLMs) can drive a real car. In the test, a model takes control of a car's steering, accelerator, and brakes (a 2022 Toyota Corolla fitted with a Comma Four device connected to the CAN bus via OBD-C) on a marked cone course roughly 130 m long in a Bay Area parking lot in California. The model receives camera frames and telemetry and issues driving commands through three MCP tools: observe(), set_motion(), and stop_now(). Progress is measured as the percentage of distance along the course centerline while staying within 4 m of it. Every run is recorded as a replayable trace, and results (progress percentage, distance, finish time, token usage) are published on a public leaderboard. The project was created in 2026, its code is released under the MIT license, and the authors emphasize they are not affiliated with or supported by comma.ai, openpilot, Toyota, or the makers of any tested models or chat applications.