DrivingBench: AI agents fail driving test, only GPT-6 Astra completes course

Researchers built DrivingBench, a test where Claude, GPT and Grok drive a Toyota Corolla around a cone-marked parking lot course via Comma hardware and OpenPilot. Eight of 11 runs failed within the first turn, Grok drove off course immediately, and only GPT-6 Astra finished the route on its second attempt for $7.74.

• 11 runs across four AI agents: eight covered under 11% of the course
• Grok drove off course on its second command, mistaking a cone gap for a gate
• GPT-6 Astra completed the course on its second attempt, costing $7.74
• One GPT model invented a cone-color rule the prompt explicitly denied