About
Contact us at team@drivingbench.com
DrivingBench is our benchmark created to test if frontier models can drive a real Toyota (through Comma 4's hardware interface and a harness/code built on top of OpenPilot) through a cone course in an empty parking lot. Each model gets up to 3 attempts in a continuous chat, issuing commands through MCP tools. A human operator sits in the driver's seat and is ready to brake at any time, and every run is recorded and published as an open trace. We hope you find this to be really cool, and we welcome suggestions for future iterations.
