METR time horizons · read 2026-11-17

How much expert work can a model do on its own?

METR measures the time horizon: how long a task is, in hours an expert human needs for it, when the AI agent finishes it 50% of the time. Below, that is mapped to a working day and week, followed by the success curve from 50% up to 99%, with a slider for the success rate.

1. Longest task a model completes at your success rate

Move the slider to set how often the model must succeed (50% to 99%). Each bar is the longest task, in expert human time, the model completes at that rate. At 50% and 80% the values are METR's own; every other rate is an estimate. 100% is not offered: on this curve the task length shrinks to zero. Dashed lines are plain arithmetic: one 8-hour working day and one 40-hour working week.

2. How long a model works on its own, from the sources

3. From 50% to 100% success: the reliability curve

Success rate as a function of task length (expert human time, log scale). Every curve starts near 100% on very short tasks and reaches 50% at the model's time horizon. The slider rate is drawn as a dashed horizontal line. The curves are derived from METR's two published points (50% and 80%) with a logistic in log task length; they are not METR's own plotted curves.

METR does not publish 90%, 95% or 99% horizons: it says measuring them would need many more, very short tasks, and results would be sensitive to methodology. Method used here: success = 1 / (1 + (t / T50)^b), where T50 is METR's 50% horizon and b = ln 4 / ln(T50 / T80) is fixed by METR's 80% horizon. The task length at success rate p is then T50 × ((1 − p) / p)^(1/b). The 90%, 95% and 99% columns are extrapolations of that fit, so treat them as estimates; they get less reliable the further p is from 80%.

What is sourced and what is assumed

Sourced: every model's 50% and 80% horizon, release date and the doubling time, from METR's data file (METR-Horizon-v1.1). Data file here: metr-time-horizons-1-1.csv.

Derived by LLMeter: the values between and beyond the 50% and 80% points, the curves, and the 90%, 95% and 99% lengths, from the two published points.

Assumption, not data: the "senior developer with 8 years of experience who knows the codebase". METR's human baseliners had about 5 years of relevant experience and little or no context for the task, and METR warns that a time horizon is not "hours of work a high-context professional does in their day-to-day job". No public source gives the speed-up for a developer who knows the codebase, so the slider starts at 1× (METR's own units). If you set it to 3×, a task METR counts as 3 hours is treated as 1 hour for that developer, so the model's horizon in that developer's time shrinks to a third.

Not covered: METR has no Claude Sonnet 5.x or Fable 5.x, and no open-weights model, so none of those can be placed on this page. This list is a selection of METR's entries from 2023 on, not all of them.

Sources: METR Time Horizons, raw data, the original paper post. Read 2026-11-17.