SenseTime published a study on 19 October that runs GPT-6-Astra across 12 simulated embodied benchmarks spanning tabletop arms, humanoids, wheeled arms and dexterous hands. Two constraints were relaxed: the model was allowed to read the training demonstrations before acting, and the step limit was raised to as much as ten times each benchmark default. Under those conditions nine RoboDojo tasks that had failed five times out of five succeeded five times out of five, and RoboDojo success reached 81.9 percent. LIBERO and SIMPLE saturated, and the report estimates near-saturation on several benchmarks at an unlimited step budget. The cost side is the counterweight: stacking bowls averaged about 131 minutes and 19 dollars of inference per task, and putting a pen into a holder averaged 263 minutes and about 55 dollars, at a rough conversion of one million tokens to one dollar. Performance stayed weakest on BEHAVIOR and on long-horizon visual navigation. Reported failure modes include not recognising that a task is finished, mishandling a change in physical state, and losing track across subtasks. A harness experiment pairs the slow planner with a fast action model and finds that checking execution results matters more than where the split is drawn. SenseTime open-sourced the SenseNova-RoboRSI harness and a 27B embodied model, and states the resulting task cost is about one fiftieth of the baseline.
Our takeThe step-limit experiment is the useful part: on these benchmarks a success rate is a function of the compute budget, so a leaderboard number without a stated budget is not comparable to anything. A flip from zero to five out of five shows those failures were search failures, not capability failures. The cost table then bounds what it means: minutes to hours and tens of dollars for a household task. The one-fiftieth cost claim is the authors own comparison and should be re-run on the released harness before it is quoted as a property of the method.