Office agents face 100 economically grounded tasks

OmegaUse-OfficeVal pairs 100 long-horizon office-suite tasks with human time and price proxies; each task takes people an average 2.32 hours, according to the open benchmark’s July 29 preprint. The evaluated frontier models were substantially faster and cheaper than people but remained below human deliverable quality, adding an outcome check to computer-use vendors’ cost-per-task claims: operators should price completed, rubric-verified work rather than inference speed or successful clicks.