METR time horizon
In short
The METR time horizon is the length of task, measured by how long a human expert takes, at which an AI agent is predicted to succeed at a chosen rate. A 50% time horizon means the prediction is success half the time.
- Also called
- Task-completion time horizon, 50% time horizon
- Topic
- Testing and benchmarks
- Level
- Technical
- First described
- The 2025 METR paper by Kwa and colleagues proposes the 50%-task-completion time horizon as a metric, and notes that using human time to put AI capability in context had been suggested earlier1.
METR estimates how long a human expert takes to finish each task, then fits a logistic curve to an AI agent's success rate against that duration. The 50% time horizon is the duration where the curve crosses a 50% chance of success. The same fit gives an 80% horizon, where the curve crosses 80%12.
The number measures task difficulty, not how long the agent runs. METR says a 50% horizon is the length of task, in human-expert time, at which it would predict that the model completes the task half the time. Agents are typically several times faster than humans on tasks they finish. The human times likely overestimate how long a professional doing equivalent work day to day would take, because the baseliners start with little prior context2.
The 2025 paper used 170 tasks from RE-Bench, HCAST and a set of shorter software tasks, built around research and software-engineering skills. On that suite the 50% horizon grew exponentially from 2019 to early 2025. The paper summarises the doubling as about every seven months, and the fitted regression puts it at 207 days. Horizons at an 80% success rate had a similar doubling time but were about four to six times shorter1.
A January 2026 update, Time Horizon 1.1, changed the suite from 170 to 228 tasks by adding 73 and removing 15. METR says the new point estimates generally fell inside the old confidence intervals, that those intervals were still very wide, and that the trend is somewhat sensitive to which tasks are included. Running some models on the old tasks in two different evaluation setups also changed their scores, so the scaffold matters as well as the task list3.
The paper says agents do worse on less structured, messier tasks, and that the suite is not the average work of researchers and software engineers1. METR's project page says the tasks are mainly software engineering, machine learning and cybersecurity, and that a horizon does not mean the agent can do every intellectual task a low-context person can do in that time2.
Example
If the fitted curve gives an agent a 50% chance of success on tasks that METR's human experts take about 30 minutes, that agent's 50% time horizon is about 30 minutes. METR says the agent will often finish a successful task faster than the human baseline2.
Why it matters
Kwa and colleagues proposed the measure because individual benchmarks saturate quickly and are hard to compare across models of very different ability1. The number still belongs to one task suite, one human baseline and one success threshold12.
Questions people ask
What is a 50% METR time horizon?
It is the human-expert duration at which a fitted curve predicts that an agent succeeds about half the time. METR reads it off success rates across tasks of different lengths. It is not the clock time the agent spends running12.
Does a two-hour time horizon mean an AI can work for two hours?
No. METR says the horizon is the length of a task in its suite, measured by human-expert time, at which it predicts a 50% chance of success. It is not the length of time an agent can act on its own2.