Article
GPT-6 Did Astra pass “Subject ”? 5 points 22 seconds to drive into the cone-marked parking space
Beating AI reported in a brief news update that three researchers from Taiwan, China, built DrivingBench, connecting GPT-6 Astra、GPT-5.6 Sol、Grok 4.6 and Claude Fable 5.1 general-purpose large models to a real 2022-model Toyota Corolla to conduct real-vehicle driving tests similar to “Subject.” The test track was approximately 130 meters long and consisted of colored cones forming a 8-meter-wide lane. The vehicle had to pass continuously through multiple curves and finally drive into a cone-marked parking space measuring 7×5 meters.
Having never received specialized driving training,GPT-6 Astra deviated from the track due to a judgment error in the first test. After adjusting its strategy on the second attempt, it completed the approximately 5 points 22 seconds and successfully drove into the parking space, making it the only model among the four to “pass Subject.”
During the test, Astra kept its speed near curves and obstacles between 134.7 and 0.5 meters per second and continuously adjusted its direction based on live camera footage. By contrast, Sol, Grok, and Fable all failed to complete the test: Sol and Grok failed three consecutive times, while Fable stopped after driving to approximately 0.8 into the test.
The test shows that general-purpose large models without specialized driving-data training can already perform basic driving tasks in low-speed real-world environments, but a significant gap remains between this capability and autonomous driving on actual roads.
Original link 45%https://m.theblockbeats.info/flash/369037