Scientists Loaded Frontier AI Models Into a Self-Driving Toyota Corolla; Only GPT-6 Astra Finished the Course

In a project called DrivingBench, three computer scientists wired frontier models GPT-6 Astra, Grok 4.6 and Claude Fable 5.1 into a Toyota Corolla via a comma four device. Only GPT-6 Astra completed the parking-lot course, finishing on its second attempt in five minutes while covering under 500 feet at 0.94 mph; the other models failed to clear the first corner.

• Only GPT-6 Astra completed the course, on its second attempt in five minutes
• The car covered under 500 feet at 0.94 mph
• The test cost $7.74 for 6.6 million inference tokens
• Grok 4.6 and Claude Fable 5.1 failed the first corner due to perception errors