Robot study finds specialist policies cut GPT-6-Astra agent's token cost

A preprint reports that a GPT-6-Astra robot agent matched the overall task success of the strongest compared specialist robot policies while substantially reducing token cost when allowed to delegate action execution to those policies on EBench. Researchers compared direct waypoint commands with a configuration that could invoke policies trained for robot actions. Direct control led on short tasks but underperformed specialist policies on long tasks, tabletop manipulation and precise contact tasks. Combining the components still fell below their best standalone performance on each task. Researchers identified poorly timed delegation, execution with fixed instructions and no subgoals, and disrupted grasps when the agent resumed control as limitations.