GPT-6.2 Astra ranked first on Artificial Analysis’s AA-Robotics v0.1 benchmark, averaging 79 out of 100 ahead of GPT-6.2 Sol at 78 and Claude Fable 5.2 at 75. The model completed six of the eight tasks and used 668 arm commands overall, requiring about 75% fewer commands than Claude Fable 5.2 for the box-cleanup task. These results suggest that the model plans its movements more effectively before acting, reducing unnecessary trial and error and hardware wear. However, each task took about 30 minutes, compared with 20 seconds for a human, while costs ranged from $2 to $16, keeping these models far from practical deployment.
Source: Read the original article


