Markdown edition

# How to read METR’s AI time-horizon chart for a work decision

> METR’s chart estimates agent success on a defined task suite. This guide shows how to record one model’s horizon, threshold and uncertainty before applying the claim to work.

By BIG CHANGE Editorial

Published: 2026-10-20T07:27:39.340Z
Updated: 2026-10-20T07:27:39.340Z
Canonical: https://bigchange.ai/blog/how-to-read-metr-ai-time-horizon-chart

![Conceptual charcoal illustration of one blank orange-tabbed index card pulled forward in a metal card file.](https://bigchange.ai/api/media/file/metr-interpretation-card-file-hero-v2.png)
AI-generated conceptual illustration by BIG CHANGE.

METR’s time-horizon chart can help an operations or research lead check a claim about an AI agent. Its plotted number is an estimate for a particular model and evaluation setup on METR’s tasks. To use it responsibly, record the exact point and then compare those tasks with the work under consideration. This guide produces a short interpretation card that another reader can check. It is based on METR’s documentation; BIG CHANGE did not run a model or test a workplace task.

## The big change

METR’s Time Horizon 1.1 chart gives a way to describe agent performance in terms of the time a human expert needs for the evaluated tasks. The 50% or 80% horizon is where a fitted success curve crosses that threshold. A team can use the chart to frame a narrower, testable question about its own work, provided it records the chart version, model, task population and uncertainty. The chart alone cannot establish that the agent will complete a particular assignment or save a team time.

## Before you open the chart

Have the actual capability claim or proposal at hand. Identify the named model, its version and, if stated, the agent setup or tools. Also write down the specific work the claim is being used to justify and what a satisfactory result would be. If the claim omits the model or chart version, request those details before accepting its number. Reading the public chart needs a browser, with no account or paid model run.

## Read one point and preserve its context

1. Open [METR’s time-horizon chart](https://metr.org/time-horizons/). Record the access date and the page’s last-updated notice. On October 20, 2026, the page said it was last updated September 8 and was no longer actively updated. Select **Time Horizon 1.1** and record that selection. The older TH1.0 series uses a different task suite; the chart is not a complete or continuously current list of model capabilities.
2. Find the exact model label used in the claim. Record any available agent setup or evaluation context. Choose **50% Success** or **80% Success** before reading the point, and note whether the overview is in log or linear scale. Copy the displayed value with its units and the view or screenshot needed to find it again. If the interactive value is inaccessible, write “unknown” in the card and leave the number out of the decision. Do not substitute a search snippet or a point from another version.
3. Read the definition under **Methodological Details**. METR estimates the time a human expert takes to complete each task, then fits a curve that predicts an agent’s chance of success as human task duration increases. The chosen horizon is the human-duration point where that curve reaches 50% or 80%. It is neither the agent’s elapsed runtime nor a maximum length of work it can undertake. It does not mean every task of that duration has the same success probability.
4. Check the task population. TH1.1’s January 29, 2026 release described 228 tasks, mainly in software engineering, machine learning and cybersecurity, drawn from RE-Bench, HCAST and shorter software tasks. They are generally self-contained, specified in advance and scored against clear criteria. METR’s human durations generally come from contracted skilled people attempting tasks with similar instructions and tools to the agents; some durations are expert estimates or quality-assurance times. A professional already familiar with a project may have more context than either participant.
5. Inspect the task-level view if it is available. Its dots represent tasks and their measured success rates; switching on **Jitter** spreads overlapping dots visually without changing performance. The method panel labels its fitted point as a **50% horizon**. Do not assume that choosing 80% in the overview changes that separate panel. METR’s FAQ illustrates why a single horizon cannot describe every task: among 90-minute-to-three-hour tasks in its GPT-5 example, about a third always succeeded in the observed runs, a third always failed and a third had mixed results. Those are the reported run patterns for that example, not permanent probabilities for those tasks or a current GPT-5 chart point.
6. Read the uncertainty notes before drawing a conclusion. METR warns that measurements above **16 human hours** are unreliable with this task suite. In the January TH1.1 release, only five of 31 tasks estimated at eight hours or longer had measured human baselines; the others used estimated times. Task selection and evaluation infrastructure can also shift estimates. The task-level fitted curve does not itself display bootstrapped confidence intervals, which are ranges estimated by resampling the evaluation results. Consult the dated release or report for intervals that apply to a particular estimate, and do not treat them as a universal error bar or as coverage of every uncertainty. METR corrected a mistake in a curve-fitting setting called regularization on March 3, 2026; numbers copied from the January release predate that correction. Its March 20 researcher note discusses fitting choices and human-duration noise, and explicitly has less review than a full publication.

## Compare the chart with the proposed work

Write down the actual task type, the context the agent and human would receive, permitted tools, the quality check, and any required interaction or approval. Compare each with METR’s self-contained, scorable tasks. A legal review, sales exchange, creative brief or project-specific engineering job does not inherit a plotted success rate because its estimated human time happens to match. Name the concrete mismatch in your own case; a missing scoring rule or reliance on internal context can change what “success” means.

Use the following card as a blank record. Fill it from the selected view and the actual proposal; do not fill a missing chart value by inference.

**Time-horizon interpretation card**

- Claim and work decision: \_\_\_
- Official chart URL, date accessed and page last updated: \_\_\_
- Exact model label, version and agent setup shown or claimed: \_\_\_
- Chart series: TH1.1; selected overview threshold: 50% or 80%; scale, filter and saved view: \_\_\_
- Displayed horizon and units, or **unknown** if unavailable: \_\_\_
- Meaning of human time and fitted threshold: \_\_\_
- Measured task domains and scoring conditions: \_\_\_
- Relevant uncertainty, date of estimate and 16-hour limit: \_\_\_
- Specific differences from our task, context, tools and success check: \_\_\_
- What this chart supports for the decision: \_\_\_
- Separately authorized local acceptance test, owner and measurable pass rule: \_\_\_
- Second reader’s source and selection check: \_\_\_

Ask a second reader to repeat the model, series and threshold selection from the same official pages, then check each factual field against the source. A complete card records a number only when the selected view supplies one. For a deployment or workflow commitment, the owner must separately authorize a test on permitted task material with an agreed success rule and human review. That test may use paid model or API capacity. The chart reading itself supplies no productivity result, safety certification or workforce forecast.

If the chart fails to load, retain “unknown” for the point and use the [TH1.1 release](https://metr.org/blog/2026-01-29-time-horizon-1-1/) for the dated method. If two estimates conflict, check their TH version, model and setup, threshold, date and whether the March correction applies; keep each value with its own label instead of averaging. Update an older card with the new citation while preserving its original access date and decision record.

## Sources & further reading

- [METR, Task-Completion Time Horizons of Frontier AI Models](https://metr.org/time-horizons/): the public chart, controls, definitions, task population, FAQ and archive notice; last updated September 8, 2026. The interactive model value must be checked in the chosen view.
- [METR, Time Horizon 1.1](https://metr.org/blog/2026-01-29-time-horizon-1-1/): the January 29 release describes the 228-task suite, long-task baselines, changes in evaluation setup and uncertainty. Its point estimates predate the March correction.
- [METR, Impact of modeling assumptions on time horizon results](https://metr.org/notes/2026-03-20-impact-of-modelling-assumptions-on-time-horizon-results/): a March 20 researcher note on the March 3 correction, fit sensitivity and uncertainty; METR labels it less reviewed.
- [Kwa and colleagues, Measuring AI Ability to Complete Long Software Tasks](https://arxiv.org/abs/2503.14499): the original 170-task methodology and limits of applying it beyond the evaluated work. The July 10, 2026 version-four author-list correction is not a TH1.1 capability release.

## Sources

- [METR task-completion time horizons](https://metr.org/time-horizons/) — Official chart and method, last updated September 8, 2026; checked October 20. Defines horizon, controls, task population, human durations, 16-hour warning and FAQ. No selected numeric point was available from the text readback.
- [METR Time Horizon 1.1](https://metr.org/blog/2026-01-29-time-horizon-1-1/) — Dated release for the 228-task suite, long-task baseline limits, evaluation changes and estimate uncertainty; January points precede the March correction.
- [Impact of modeling assumptions on time horizon results](https://metr.org/notes/2026-03-20-impact-of-modelling-assumptions-on-time-horizon-results/) — Less-reviewed METR researcher note on the March 3 fitting correction, alternative fits and human-duration noise; not a new capability release.
- [Measuring AI Ability to Complete Long Software Tasks](https://arxiv.org/abs/2503.14499) — Original 170-task method and limits. Version four corrected its author list on July 10; that date is not a TH1.1 release or new measurement.