GPT-6 Astra Passes 'Subject 2'? Parks into Cone Space in 5 Minutes 22 Seconds

Beating AI News Flash: Three researchers from the China Taiwan region built DrivingBench, connecting general-purpose large models such as GPT-6 Astra, GPT-5.6 Sol, Grok 4.6, and Claude Fable 5.1 to a real 2022 Toyota Corolla for a real-vehicle driving test similar to China's "Subject 2" driving exam. The test track is approximately 130 meters long, with colored cones forming an 8-meter-wide passage. The vehicle must navigate through multiple consecutive curves and finally park into a 7×5 meter cone-marked parking space.

GPT-6 Astra, which had never received specialized driving training, deviated from the track in its first test due to a judgment error, but after adjusting its strategy on the second attempt, it completed the approximately 134.7-meter course in 5 minutes and 22 seconds and successfully parked into the space, becoming the only model among the four to "pass Subject 2."

During the test, Astra controlled its speed at 0.5 to 0.8 meters per second near curves and obstacles, continuously adjusting its direction based on real-time camera footage. In comparison, Sol, Grok, and Fable all failed to complete the test, with Sol and Grok failing three consecutive times, and Fable stopping at approximately 45% of the course on its third attempt.

This test demonstrates that general-purpose large models without specialized driving data training are already capable of completing basic driving tasks in low-speed real-world environments, but there remains a clear gap before achieving autonomous driving on actual roads.