METR defines a task-completion time horizon as the duration, measured by human expert completion time, at which an agent is predicted to succeed with a specified reliability. The 50% horizon is the point where the fitted model predicts success half the time. This definition matters because it separates task difficulty from elapsed autonomous runtime: the measure is not how long an agent can work continuously.
METR fits a logistic curve to predicted success against human-expert task duration, then reads the duration at the specified success probability. METR says its current suite draws primarily on software engineering, machine learning, and cybersecurity tasks, designed to be self-contained, well-specified, and automatically evaluable. METR reports a predicted 50% time horizon of 17 hours for GPT-4. It also says measurements above 16 hours are unreliable with the current task suite. The task set and estimator therefore make a capability comparison possible, but they do not turn a benchmark duration into a claim about sustained independent work.
Builders can use the measure to reason about delegation of bounded tasks, but should not read it as a project-autonomy estimate. The public page presents a current comparison of frontier agents alongside the metric definition and caveats, making it usable as bounded capability evidence. The evidence covers METR’s largely self-contained software engineering, machine learning, and cybersecurity task suite; it does not establish performance on context-rich professional work or autonomous research end to end. Watch whether future evaluations broaden task coverage and report reliable measurements beyond the current upper range. No prior New Runtime record was supplied for this measurement page, so this is a distinct analytical record.