Did GPT-6 Astra Pass "Subject 2"? It Pulled Into the Cone-Marked Parking Space in 5 Minutes and 22 Seconds.

Beating AI Express (Dongcha) reports that three researchers from Taiwan, China have built DrivingBench, integrating general large models including GPT-6 Astra, GPT-5.6 Sol, Grok 4.6 and Claude Fable 5.1 into a real 2022 Toyota Corolla to conduct real vehicle driving tests similar to China’s driving test Subject 2. The test track is approximately 130 meters long, with an 8-meter-wide channel formed by colored traffic cones. The vehicle needs to navigate multiple curves before parking into a 7×5-meter cone-marked space. GPT-6 Astra, which had no prior specialized driving training, deviated from the track due to misjudgment in its first test. After adjusting its strategy in the second attempt, it completed the roughly 134.7-meter course in 5 minutes and 22 seconds and successfully parked, becoming the only one of the four models to "pass Subject 2". During the test, Astra controlled its speed at 0.5 to 0.8 meters per second near curves and obstacles, adjusting direction continuously based on real-time camera footage. In contrast, Sol, Grok and Fable all failed to finish the test: Sol and Grok failed three consecutive times, while Fable stopped after reaching around 45% of the course on its third run. The test shows that general large models not trained on specialized driving data can now complete basic driving tasks in low-speed real-world environments, though they still have a significant gap before achieving actual road autonomous driving.