METR’s AI time horizon estimates the length of a task, measured by the time it takes a human expert, at which an AI agent is predicted to succeed at a chosen rate. A 50% horizon is not a clock measuring how long the agent can work without interruption.
The metric is useful for comparing performance on a particular suite of tasks. It is not a general measure of whether an agent can do every job that takes a person the same amount of time. This explanation uses METR’s published methodology, checked October 31, 2026.
On this page
How METR calculates a time horizon
METR first estimates how long a human expert takes to complete each task. It then runs an AI agent on tasks of different human-measured lengths and fits a curve relating task duration to the probability of success.
The point where that curve crosses 50% is the estimated 50% time horizon. METR also reports an 80% horizon. The number describes a predicted success rate for the task distribution and setup used in the evaluation, rather than a guarantee for each individual task.
For example, a hypothetical two-hour horizon means the fitted curve predicts 50% success at a task length of two human expert hours. It does not mean the agent itself spends two hours on the task, or that every two-hour task has exactly the same chance of success.
What counts as a task?
The Time Horizon 1.1 suite draws tasks from RE-Bench, HCAST and shorter software tasks. METR says they primarily involve software engineering, machine learning and cybersecurity. They are designed as self-contained, well-specified units with clear criteria that can be scored automatically.
METR uses successful human attempts to estimate duration for most tasks, taking a geometric mean. Where reliable measurements are unavailable, it uses expert estimates or quality-assurance completion times. Its researchers caution that these estimates can differ from the time taken by someone already familiar with a real project.
An agent is evaluated with a model and a scaffold: the tools and interaction loop that let it work. That setup matters when comparing results. METR describes running multiple independent attempts and reviewing potential reward hacks before analysis.
What the number does not show
A longer horizon does not establish that a model can perform all work of that duration. METR’s suite concentrates on selected technical tasks, while jobs often involve tacit knowledge, interactions with people and less tidy definitions of success. Performance can vary by task even at the same measured length.
The metric is also not a measure of uninterrupted running time. The duration on the horizontal axis is a human baseline. An agent may complete a task in less or more time depending on its setup and the task.
METR warns that estimates above 16 hours are unreliable with its current task suite. Its public time-horizon page was last updated September 8, 2026 and says it is no longer actively updated. Readers should check the evaluation date and method before comparing a number with a newer report.
How to read a reported result
Look for the reliability threshold first: a 50% horizon and an 80% horizon answer different questions. Then check the task suite, agent scaffold, evaluation date and uncertainty. A model name alone does not describe the tools or conditions under which the measurement was obtained.
The original 2025 research article explains the curve-fitting approach and its limitations. Later evaluations can use a different suite or setup, so changes in methodology matter when interpreting a trend.
