A Simple Measure For AI: How Long It Can Work On A Task
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: A Simple Measure For AI: How Long It Can Work On A Task on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

METR’s public benchmark data, as reported by ThorstenMeyerAI.com on Nov. 18, 2026, puts an early Claude Mythos Preview model’s 80% task horizon at 185.9 minutes—about three hours of human task time. The measure estimates how long a skilled person would take to complete tasks that an AI agent can finish at a specified success rate; it does not mean the AI works for that duration or will succeed on every task.

A public METR benchmark puts an early version of Anthropic’s Claude Mythos Preview at an 80% task horizon of 185.9 minutes, or about three hours of work as timed for a skilled human. The figure, reported by ThorstenMeyerAI.com on Nov. 18, 2026, offers teams a way to estimate which tasks AI agents may be able to complete reliably, but it is a benchmark result—not a guarantee for any individual job.

METR, an independent research nonprofit, defines a task horizon by the length of tasks, measured in human expert working time, that an AI agent is predicted to complete at a chosen reliability level. At the 80% threshold, the estimate means success on roughly four out of five attempts across the benchmark’s tasks at that length. It does not mean the model itself spends three hours working: METR says the measure concerns task difficulty, and agents can finish successful tasks faster than people.

The report says METR’s Time Horizon 1.1 suite contains 228 tasks, spanning software engineering, machine learning and cybersecurity work. Skilled professionals first complete and time the tasks. The AI then receives six independent runs per task; transcripts are checked for cheating, which counts as failure. METR fits the relationship between task length and success rate to estimate where it crosses 80% or 50%.

According to the cited results file, the 80% horizon rose from 0.89 minutes for GPT-4 in March 2023 to 29.98 minutes for o3 in April 2025 and 66 minutes for GPT-5.2 in December 2025. The early Mythos Preview entry is dated April 7, 2026. Its reported confidence interval is wide, from 97.3 to 398.5 minutes, so the point estimate should not be read as exact. The underlying data were retrieved Nov. 18, 2026, but the report says METR’s public results page was last updated Sept. 8 and is no longer actively updated.

At a glance
reportWhen: Report published Nov. 18, 2026; the cit…
The developmentA Nov. 18 report using METR benchmark data identifies an early Claude Mythos Preview model as having an 80% task horizon of 185.9 minutes, while highlighting limits in reliability and how well benchmark tasks represent real work.

Using Task Length to Plan AI Work

The horizon gives managers a more practical question than whether AI can perform a broad category of work: how long would a skilled person need for this task, and how often does the agent complete comparable work? That framing can help teams choose tasks, divide larger assignments into smaller pieces, and budget time for review. It also makes clear why a successful demonstration on a short, well-defined task does not establish that an agent can handle a long, ambiguous assignment.

The 80% figure still leaves about one failure in five at the stated threshold. The report recommends treating verification as part of the work, giving agents tasks with clear acceptance checks, and measuring results on a team’s own tasks. This matters where a missed defect could create operational or security problems: a benchmark score cannot replace testing or accountable human review.

There are reasons to be cautious about applying the estimate directly to workplace productivity. ThorstenMeyerAI.com cites a 2025 METR trial in which experienced developers were 19% slower with AI while believing they were 20% faster. It also cites a separate study where reviewers caught 31% of errors that required applying the task’s rules. These are findings reported by the source, not proof that every team or tool will have the same results.

Products Worth Considering

Logitech MX Creative Console with 9 Customizable Keys, Graphite

Logitech MX Creative Console with 9 Customizable Keys, Graphite

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

As an affiliate, we earn on qualifying purchases.

How METR Measures Task Horizons

The measure uses human completion time as its unit, rather than the amount of time an AI tool runs. Researchers compare success across tasks of different lengths, then estimate the point where performance reaches a selected reliability level. METR’s 80% horizon is intended as a more dependable planning reference than a 50% horizon, which represents success about half the time.

The reliability threshold changes the result substantially. For the cited Mythos Preview entry, the reported 50% horizon is 1,044.8 minutes, or about 17.4 hours, compared with 185.9 minutes at 80%. METR cautions that measurements above 16 hours are unreliable with its current task suite, so the longer estimate is especially uncertain. The source also reports that messier tasks reduced the measured horizon by nearly a factor of three in one METR comparison, a warning that clean benchmark assignments may not match routine work.

The source says its own trend fits imply that an eight-hour workday at the 80% level could fall between September 2026 and November 2027, depending on which years are included. That is a conditional forecast, not a measured result or a confirmed timeline. It also reports that nothing after April 2026 had been measured publicly in the cited record.

Limits in the Public Benchmark

The 185.9-minute figure is a fitted estimate with a broad confidence interval, and the public record described in the report has not been actively updated since Sept. 8, 2026. It is therefore unclear whether newer models or evaluations would change the current picture. The report does not provide enough detail to establish how the score transfers to every workplace, task type, or operating environment.

METR’s benchmark tasks are described as clean and well specified, while real assignments may contain incomplete instructions, shifting requirements, dependencies, or hidden constraints. The source reports a near-threefold drop in horizon for messier tasks in one comparison, but does not supply enough detail here to treat that reduction as a universal adjustment. Nor does an 80% aggregate success estimate tell a team which particular attempt will fail.

The longer 50% horizon is also uncertain: it exceeds the 16-hour range METR says its task suite can measure reliably. The proposed timeline for reaching an eight-hour 80% horizon is a modelled extrapolation, not a public benchmark result. Both the estimate and the forecast should be kept separate from confirmed task-level performance.

Measuring Results on Team Tasks

For teams applying the measure, the report recommends starting with work that can be divided into hour-scale tasks with clear acceptance checks, then tracking completion and review effort. Teams can compare models on their own work, keep a record of failures and corrections, and repeat evaluations periodically; the source suggests re-measuring each quarter. Those steps would show whether external benchmark results transfer to a particular team’s code, processes, and risk tolerances.

The next public milestone is a clearer, updated evaluation record: the cited METR page is no longer actively updated, and the report says public measurements after April 2026 were unavailable. Until updated data or task-specific evaluations appear, the three-hour figure is best treated as a dated benchmark estimate, not a forecast of guaranteed workplace capacity.

Key Questions

What does an AI task horizon measure?

It estimates the length of a task, measured by how long a skilled human takes, at which an AI agent reaches a specified success rate. It does not measure how long the AI itself works.

What does an 80% horizon mean?

It means the fitted estimate is for tasks of that human-time length where the agent succeeds about four times in five across benchmark attempts. It does not promise success on a specific task.

Which model has the reported three-hour horizon?

The cited METR results file lists an early Claude Mythos Preview, dated April 7, 2026, with an 80% horizon of 185.9 minutes. The reported confidence interval ranges from 97.3 to 398.5 minutes.

Can teams use the figure to estimate workplace productivity?

Only as a starting point. METR’s benchmark tasks may be cleaner than real work, and success rates do not account for every review or correction cost. Teams need evaluations on their own tasks before using the score for planning.

Is an eight-hour AI workday confirmed?

No. The source reports a fitted forecast ranging from September 2026 to November 2027, depending on the years used. It is an extrapolation, not a measured or confirmed result.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

Mistral Forge: Owning the Model, Not Just Renting the API

Mistral’s Forge offers organizations the ability to own and operate their own AI models, moving beyond API rental. This development impacts enterprise AI sovereignty and data control.

Mistral Large 4 Trails The Leading Edge Of AI

Mistral Large 4 is available through a preview API, but benchmark results put it below leading US and Chinese models. Its weights are due later in October.

A Closer Look At AI’s Potential For Recursive Self-Improvement

Anthropic and METR report AI-assisted gains in AI research, but fully autonomous recursive self-improvement has not been demonstrated.

Tech Operations And The Stability Of URIs: A Historical Insight

Explores the history and significance of URI stability in technology operations, highlighting lessons from the 1998 ‘Cool URIs Don’t Change’ principle.