Autonomous completion of long tasks

The measure is the length, in skilled-human time, of a task a model finishes half the time; at 80 percent reliability the horizon is several times shorter, and the tasks are self-contained software and research problems with clear success criteria. METR's 2026 revision put the recent doubling faster, around four months, and reported horizons above ten hours for the newest models, with a warning that measurements beyond sixteen hours are unreliable on the current suite. Weeks-long work with shifting goals, other people, and consequences that cannot be reset is what remains; the steepness of the curve is why this entry is contested rather than held.