An anonymously published simulation evaluation report on GitHub shows that OpenAI's newly released GPT-6 Astra model demonstrates zero-shot generalization across multiple robotic manipulation tasks without robot-specific adaptation. When combined with the robot vision-language-action model π0.5 into a hybrid architecture, it achieved an average score of 62.6 on the RoboDojo benchmark, 64% higher than the second-place score of 38.26. The report's authors come from the Galaxy General robotics team, with the first author being a PhD student under Professor Wang He, the company's founder and CTO. GPT-6 Astra demonstrated robotic policy capabilities including proposing actions based on task semantics, correcting target alignment errors, performing non-grasping operations, and reasoning and recovering based on execution feedback, while weaknesses were exposed in tasks requiring fine contact and stable control, such as block stacking and Mahjong tile manipulation. The team designed Direct mode, in which π0.5 was not run and GPT-6 Astra directly viewed three camera feeds and robot states, outputting dual-arm end-effector poses and gripper open/close commands, executing 1 to 5 control steps at a time, and Hybrid mode, in which π0.5 first generated 50 candidate action steps at each decision point and GPT-6 Astra viewed the same visual inputs, execution history, and the dual-arm trajectories corresponding to the candidate actions, then decided whether to adopt π0.5's output or perform its own end-effector pose