From minutes to hours of work

METR: software, ML and cybersecurity tasks. Italic annotations are METR’s example tasks, positioned by their supplied human duration; they are not model evaluations. Solid points are fitted time horizons; bars show the supplied confidence bounds. Above 16 hours, METR warns that this task suite gives unreliable measurements. METR TH 1.1; illustration inspired by Luke Leisher. Hollow diamonds use the separate upper-right agent wall-time axis, not the human-time axis. These are different metrics, not additional measured 50% horizons. Navier–Stokes: 88 h with ~10,000 concurrent agents; Hugging Face: 107.77 h recovered campaign (July 9–13). Parallel agent count is not multiplied into the y-value.