Assessing AI’s Reach: Measure The Tasks It Can Do In Hours
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Assessing AI’s Reach: Measure The Tasks It Can Do In Hours on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

A November 18, 2026 report explains METR’s task-duration benchmark, which measures how long a skilled human would need for work an AI agent completes at a specified success rate. Its cited record is 185.9 minutes at 80% success for an early Claude Mythos Preview, but the benchmark is limited and the public results stop at April 2026.

A November 18 report from ThorstenMeyerAI.com says the strongest publicly listed result in METR’s task-duration benchmark is 185.9 minutes at an 80% success rate, or about three hours of work as measured by skilled human completion time. The result, for an early Claude Mythos Preview dated April 7, 2026, measures task difficulty and reliability—not the time the AI itself spends working—and does not establish how the system performs on ordinary workplace assignments.

METR, an independent research nonprofit, evaluates AI agents on tasks timed by skilled professionals in fields including software engineering, machine learning and cybersecurity. Its current cited suite, Time Horizon 1.1, contains 228 tasks. The report says METR has professionals with an average of about five years of relevant experience time the work, then runs agents on tasks multiple times and assesses whether they succeed. The benchmark’s time horizon estimates the human task duration at which an agent reaches a chosen success rate.

The report focuses on the 80% horizon: the task length at which the fitted success curve reaches four successful attempts out of five. METR’s results file lists the Mythos Preview at 185.9 minutes, with a confidence interval from 97.3 to 398.5 minutes. For comparison, the report cites GPT-4 at 0.89 minutes in March 2023, o3 at 29.98 minutes in April 2025, and GPT-5.2 at 66 minutes in December 2025. These are benchmark results, not guarantees for individual jobs.

Reliability changes the estimate substantially. At a 50% success threshold, the same Mythos Preview result is listed as 1,044.8 minutes, or about 17.4 hours. The report notes that METR warns measurements above 16 hours are unreliable with its current task suite. It also says METR’s public results page is no longer actively updated and was last updated September 8, 2026, leaving the cited record dependent on measurements through April.

At a glance
reportWhen: Report published November 18, 2026; cit…
The developmentA November 18 report uses METR’s task-duration benchmark to assess the length and reliability of work AI agents can complete.

How Teams Can Use the Horizon

The benchmark frames AI evaluation around task duration and completion rates. That framing can help teams select work for trials, divide projects and estimate the human review required. A three-hour 80% horizon indicates performance on benchmark tasks at that level of human effort; it does not establish that an agent can independently manage a three-hour block of a worker’s day.

The report recommends breaking work into short, clearly specified pieces, starting with tasks that have testable acceptance criteria, and planning time for review. It also advises teams to compare models against their own work and repeat those measurements quarterly. Local testing can account for differences between benchmark tasks and a company’s code, data, procedures and tolerance for mistakes.

Oversight has a measurable cost. The report cites a 2025 METR trial in which experienced developers using AI were 19% slower, although they believed they were 20% faster. It also cites a separate review finding that reviewers caught only 31% of errors requiring task-specific rules to identify. These findings do not determine AI’s productivity effect across workplaces, but they show why task-completion rates alone do not measure workplace value.

Products Worth Considering

Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera – Midnight

Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera – Midnight

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

As an affiliate, we earn on qualifying purchases.

What METR’s Clock Measures

METR’s measure is based on human working time, not the AI’s runtime. A task that takes an expert two hours may be completed by an agent faster when it succeeds; the horizon instead describes the length of human-rated tasks at which success falls to a chosen level. The report says agents are typically several times faster than humans on tasks they complete, but speed does not remove the need to check the result.

To calculate a horizon, METR times human experts, attempts each task with an AI agent in multiple independent runs, and plots success against task length. The report says the current suite uses six runs per task and treats cheating detected in transcripts as failure. The fitted curve’s crossing point provides the estimate at either 80% or 50% reliability.

The benchmark’s controlled tasks are not a complete proxy for workplace conditions. The report says that in one METR comparison, messier tasks reduced the measured horizon by almost a factor of three. It also describes a trend of horizons doubling every three to seven months, depending on the period and fit used, while noting that this is a projection rather than a dependable schedule. Its estimate for reaching a full eight-hour workday at 80% ranges from September 2026 to November 2027 across different fits.

“the task duration (measured by human expert completion time) at which an AI agent is predicted to succeed with a given level of reliability.”

— METR, as quoted in the report

Limits of the Public Benchmark

The cited record has a wide confidence interval, spanning roughly 1.6 to 6.6 hours, and the report says the benchmark is not actively updated. The cited public data does not show how leading models released after April 2026 would score under the same test. The report’s statement that no later result has been measured publicly refers to the source’s information as of November 18, 2026.

It is also unclear how closely the suite reflects routine work across industries. The report says task messiness can sharply lower the measured horizon, and results from a limited set of technical tasks cannot establish performance on every team’s workflows. An 80% aggregate estimate does not promise success on any particular task; failures remain part of the result.

The report cites studies about developer speed and error detection, but those findings do not establish the net effect of AI across different roles, tools or review processes. The evidence presented does not establish whether productivity gains outweigh setup, checking and rework costs for a specific organization.

Testing Against Real Work

Employers can test AI on their own representative tasks, using clear success criteria and recording both completion quality and review time. Teams can begin with bounded work that is easy to verify, then compare results across models and revisit the measurements as tools change.

Updated public METR results would show whether models released after April 2026 extend the measured horizon or change its uncertainty. Until then, the three-hour figure is a dated benchmark estimate, not a forecast of when AI will reliably take over a full workday.

Key Questions

What does a three-hour AI task horizon mean?

It means the benchmark estimates that an agent reaches a specified success rate on tasks that take skilled humans about three hours. It does not mean the agent spends three hours working or can reliably perform every three-hour job.

Which model has the report’s cited record?

The report cites an early Claude Mythos Preview, dated April 7, 2026 in METR’s results file, with an 80% horizon of 185.9 minutes. The confidence interval is wide, and the report says the public benchmark is no longer actively updated.

Why is the 50% result much longer than the 80% result?

A lower success threshold allows longer tasks into the estimate. The cited Mythos Preview result is about 17.4 hours at 50% success, compared with about 3.1 hours at 80%; METR cautions that measurements above 16 hours are unreliable with its current task suite.

Can companies use the benchmark to predict productivity gains?

Not on its own. It can help identify tasks for controlled trials, but benchmark performance does not account fully for company-specific work, checking, errors or rework. Teams need to measure outcomes on their own tasks.

Source: ThorstenMeyerAI.com

You May Also Like

Get Ready For The Game With New Football Features In Search

Google Search introduces new football features to help fans prepare for upcoming games, with details still emerging about the full scope and capabilities.

Why the Worst AI Manager Still Scores 26: Inside a Benchmark That Refuses to Grade on a Curve

A benchmark where doing nothing scores 26, one breach of trust caps your grade, and only 2 of 5 AI models closed a deal they’d already earned.

Fine-tuning A 350M Model For Better Structured Outputs In 100 GRPO Steps

Researchers demonstrate a method to fine-tune a 350M parameter AI model for improved structured outputs using only 100 gradient steps, promising faster adaptation.

How To Get Structured JSON Fields From A Document With Dharma

Hugging Face says its 2-billion-parameter model converts document images into JSON, with benchmark and cost advantages that still need independent testing.