AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Understanding What AI Can Do Today By Measuring Time on ThorstenMeyerAI.com

TL;DR

METR’s public benchmark records an 80% task horizon of 185.9 minutes for an early Claude Mythos Preview released in April 2026. The measure estimates how long skilled people take to complete tasks that an AI agent can finish successfully four times out of five—not how long the AI itself works. Benchmark conditions, task reliability and the absence of public measurements after April 2026 limit what the figure says about workplace performance today.

A November 18 report from ThorstenMeyerAI.com uses METR’s task-duration benchmark to measure what AI agents can do today, pointing to a public record of 185.9 minutes at an 80% success rate for an early Claude Mythos Preview. The figure describes tasks that take skilled human workers about three hours; it does not mean the model works for three hours or succeeds on every comparable job.

METR, an independent research nonprofit, measures task length using the working time of skilled human experts. It then has AI agents attempt tasks several times and estimates how the probability of success changes as tasks get longer. The 80% time horizon marks the task duration at which the fitted success curve reaches four successful attempts in five. METR also reports a 50% horizon, a less demanding reliability threshold.

In the results file cited by the report, the April 7, 2026, Claude Mythos Preview entry has an 80% horizon of 185.9 minutes. Its reported confidence interval spans 97.3 to 398.5 minutes, or roughly 1.6 to 6.6 hours. The same model’s 50% horizon is listed as 1,044.8 minutes, about 17.4 hours, but METR cautions that measurements above 16 hours are unreliable with its current task suite.

The benchmark’s earlier entries show the 80% measure rising 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. These are results from METR’s benchmark, not a direct measure of every product or work setting. The guide says the public results page is no longer actively updated and that its latest update was September 8, 2026.

At a glance
reportWhen: Published November 18, 2026; the latest…
The developmentA guide published November 18, 2026, uses METR’s public benchmark to explain how AI capability can be measured by the length of tasks models complete reliably.

How Task Length Shapes AI Planning

The measure offers teams a more practical planning question than whether AI can perform a task at all: how long does the task take a skilled person, and how reliably does the model finish it? A three-hour horizon can help frame experiments with work that has a clear scope and a way to check the result. It should not be read as a guarantee that any three-hour job will succeed.

The report recommends breaking work into short, testable pieces, starting with tasks that have clear acceptance checks, and budgeting time for review. That matters because an 80% success rate still means one failure in five across the benchmark’s attempts. Teams may face different error rates on their own work, so the benchmark cannot replace local testing.

The report also cites a METR 2025 trial in which experienced developers were 19% slower while using AI, despite believing they were 20% faster. It references a separate study where reviewers caught 31% of errors that required knowledge of task-specific rules to identify. Those findings are reported as evidence that tool use and review can affect productivity; they do not establish that every team or AI system will have the same results.

Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black

Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black

As an affiliate, we earn on qualifying purchases.

How METR Builds Its Measure

METR’s current Time Horizon 1.1 suite contains 228 tasks, according to the source report. Tasks cover fields including software engineering, machine learning and cybersecurity. METR contracts professionals with an average of about five years of relevant experience to complete tasks and records their working time, giving each task a human-duration estimate.

AI agents then receive six independent runs per task. METR checks transcripts for cheating and counts it as failure. It plots success against human task duration and fits a curve; the point where that curve reaches 80% is the 80% horizon. The result is a benchmark estimate across tasks, not a timer showing how long an AI spends working. METR says agents typically complete tasks they solve several times faster than humans.

The source report also notes a mismatch between benchmark conditions and ordinary work. METR tasks are described as clean and well specified; one comparison found that messier tasks reduced the horizon by almost a factor of three. That comparison is a warning about generalizing the score, not a universal conversion factor for workplace projects.

Limits of the Three-Hour Estimate

The public record has a wide confidence interval, and METR says its current task suite cannot reliably measure horizons above 16 hours. The 50% result of about 17.4 hours should be treated with particular caution. Neither score is a promise about an individual task, since the figures describe fitted success rates across benchmark attempts.

It is also unclear how closely these results predict performance on a particular company’s code, documentation, or operational processes. The source report says less-structured tasks can reduce the measured horizon, but does not provide a single adjustment that applies across industries. METR’s public results page is no longer actively updated, and the cited material contains no public measurements later than April 2026.

The report gives a projection that an eight-hour workday at 80% reliability could fall between September 2026 and November 2027, depending on which years of the trend are used. That is the report’s own fit, not a confirmed milestone: it depends on the trend continuing, and the cited public record does not establish that the benchmark has since reached an eight-hour horizon.

Testing Models on Real Work

For teams, the next step suggested by the report is to select bounded tasks with clear pass-or-fail checks, measure how often the AI completes them, and include human review in the time and cost calculation. Repeating those tests on a team’s own work every quarter would help show whether results change as models or workflows change. METR’s public task horizon can provide a reference point, but it does not settle that local question.

Further public benchmark updates would clarify whether the measured horizon has moved since April 2026 and whether newer results remain comparable with Time Horizon 1.1. Until such measurements are available, the three-hour figure is best treated as the latest cited benchmark result, with substantial uncertainty around its range and its fit to less structured work.

Key Questions

What does an AI time horizon measure?

It estimates the duration, in skilled human working time, of tasks at which an AI agent reaches a stated success rate. METR reports results at thresholds including 80% and 50%.

Does a three-hour horizon mean the AI works for three hours?

No. It refers to tasks that take a skilled person about three hours. The measure concerns task difficulty and estimated reliability, not the time the AI spends completing a task.

What is the latest public 80% result cited in the report?

The report cites 185.9 minutes for an early Claude Mythos Preview dated April 7, 2026, in METR’s results file. The confidence interval runs from 97.3 to 398.5 minutes.

Can companies use the result to predict AI productivity?

Not by itself. METR’s benchmark uses specified tasks and measured human completion times. The report says messier tasks can shorten the horizon, so teams need to test their own work and account for checking and failures.

Has METR published a later public result?

The source report says METR’s public results page is no longer actively updated and cites no public measurement later than April 2026. Whether a newer public benchmark result has since been released is not established by the material provided.

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

India And AWS Feature In Anthropic’s Claude AI Inference Expansion

A Times of India headline reports that Claude inference is available in India through AWS, but launch and data-handling details remain unconfirmed.

Will The Lowest Temperature In Seoul Be 21°C On July 17?

Speculation surrounds Seoul’s weather forecast for July 17, with some predicting a low of 21°C. Experts clarify what is confirmed and what remains uncertain.

We Tried Weight-Space Programming At Home: Claude Edited A Local Model, GPT-6 Astra Checked The Work, And The Legal Score Stayed At Zero

A weight edit helped a local model finish 10 of 10 Lexassay tasks, but its legal score stayed at zero and its answers contained citation errors.

The Bubble Is Not in Valuations: It’s in the Productivity Gap

Analysis of the emerging ‚expectation bubble‘ in AI, where projected productivity gains are not yet measurable, risking long-term economic impacts.