AI At Work: Measure Its Capabilities By Task Duration
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: AI At Work: Measure Its Capabilities By Task Duration 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 measures AI capability by the amount of skilled human work a task requires and the rate at which an AI agent completes tasks of that length. Its latest public record, for an early Claude Mythos Preview in April 2026, is 185.9 minutes at an 80% success rate, but the benchmark has important limits and has not been publicly updated since.

A report published November 18 by ThorstenMeyerAI.com says the strongest model in METR’s public task-duration results completed tasks requiring about three hours of skilled human work with an estimated 80% success rate. The measure offers teams a way to judge which work AI may handle, but it is not a guarantee for any particular job: the cited record is an early Claude Mythos Preview result from April 2026, and METR’s public results page has not been actively updated since September.

METR, an independent research nonprofit, defines its time horizon as the length of a task, measured by human expert completion time, at which an AI agent is predicted to succeed at a specified reliability level. The report uses the 80% horizon as a planning benchmark: a model is estimated to complete four out of five tasks of that duration, rather than every task reliably.

According to METR’s Time Horizon 1.1 results file, the early Claude Mythos Preview reached 185.9 minutes at 80% reliability, based on a result dated April 7, 2026. The report compares that with 0.89 minutes for GPT-4 in March 2023, 29.98 minutes for o3 in April 2025 and 66 minutes for GPT-5.2 in December 2025. Those figures are benchmark results, not a claim that the model spends three hours working: METR says the measure reflects task difficulty, and that agents can finish successful tasks faster than people.

The report says METR’s current Time Horizon 1.1 suite includes 228 tasks, spanning seconds to more than eight hours of skilled human work. METR uses professionals in software engineering, machine learning and cybersecurity to establish task times, then runs each task six times with an AI agent. It checks transcripts for cheating, which counts as failure, and estimates the duration at which the success curve crosses the selected reliability threshold.

At a glance
reportWhen: Reported November 18, 2026; the cited p…
The developmentA November 18, 2026 report from ThorstenMeyerAI.com explains METR’s task-duration measure and reports a public benchmark record of 185.9 minutes at 80% reliability.

How Teams Can Use the Horizon

The measure shifts workplace planning away from broad questions such as whether AI can “do” a job and toward which tasks it can complete at a useful reliability level. For managers, that can help identify work suitable for a supervised trial: bounded tasks that take hours, have clear acceptance criteria and can be checked against tests or other evidence.

The report recommends breaking work into short, testable pieces, starting with tasks where success is easy to verify and budgeting time for review. It also advises teams to compare models on their own work and repeat those evaluations quarterly. These are recommendations from the report, not findings that every organization will see the same results.

That distinction matters because a benchmark success rate leaves failures to manage. At an 80% rate, an estimated one in five attempts fails. The report cites a 2025 METR trial in which experienced developers working with AI took 19% longer, despite believing they were 20% faster. It also cites a separate study where reviewers caught 31% of errors that required task-specific rules to identify. Those findings caution against treating AI assistance as an automatic productivity gain or assuming that review will catch every mistake.

Breville Barista Express Espresso Machine with Built-In Grinder, Brushed Stainless Steel

Breville Barista Express Espresso Machine with Built-In Grinder, Brushed Stainless Steel

Shark PowerPro Reveal Plus Cordless Stick Vacuum, Charcoal (IZ376H)

Shark PowerPro Reveal Plus Cordless Stick Vacuum, Charcoal (IZ376H)

As an affiliate, we earn on qualifying purchases.

What METR’s Benchmark Measures

The time horizon is one way to summarize performance across tasks of different lengths. METR plots success against the amount of human work needed: short tasks tend to have higher success rates, while longer tasks tend to have lower ones. The point where the estimated curve crosses 80% success is the 80% horizon; the corresponding 50% point is a less reliable threshold for operational planning.

The report’s headline figure comes with substantial uncertainty. It gives a confidence interval of 97.3 to 398.5 minutes for the Mythos Preview result, equivalent to roughly 1.6 to 6.6 hours. It also reports that the model’s 50% horizon is 1,044.8 minutes, or about 17.4 hours, but notes METR’s warning that estimates above 16 hours are unreliable with the current task suite. That longer estimate should not be read as a dependable workday capability.

Benchmark conditions matter, too. The report says METR tasks are relatively clean and well specified, and cites a comparison in which messier tasks reduced the measured horizon by nearly a factor of three. The public record therefore describes results on a particular suite, not a universal measure of performance across company systems, ambiguous requests or routine office work.

“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 ThorstenMeyerAI.com report

Limits of the Public Record

The best-known public result is dated April 7, 2026, and the report says METR’s results page was last updated September 8, 2026, with no active updates. It is not clear from the supplied material whether newer models have been evaluated internally, or when METR may publish another public measurement.

The confidence interval is wide, and the result depends on a task suite that may not represent messy workplace assignments. The supplied report also describes forecasts that an eight-hour workday horizon at 80% could arrive between September 2026 and November 2027, depending on the years used in the trend fit. That is a forecast, not an observed result; the report says the measured line had doubled every three to seven months if the trend held.

It also remains unclear how well benchmark rates predict performance on any particular organization’s code, policies or review process. The supplied material does not establish that a model will achieve the same success rate on a company’s own tasks, or that the cited developer and reviewer studies apply to every team.

Testing AI on Company Work

For organizations considering AI agents, the next practical step described in the report is to test them against their own representative tasks, with clear pass criteria and a record of failures. Teams can begin with bounded, verifiable work, compare tools by task rather than by general reputation and include human review in time estimates.

METR’s public benchmark may provide a further reference if it is updated, but the supplied material gives no date for a new release. Until then, the 185.9-minute figure should be treated as a dated estimate from a specific benchmark, with a wide confidence interval—not as a forecast of guaranteed capability or a substitute for local testing.

Key Questions

What does a three-hour AI time horizon mean?

It means the model is estimated to succeed at the chosen reliability level on tasks that skilled humans take about three hours to complete. It does not mean the AI itself works for three hours or will finish every such task.

Which model has the highest result in the cited public data?

The report identifies an early Claude Mythos Preview result dated April 7, 2026, with an 80% time horizon of 185.9 minutes. The report gives a wide confidence interval of 97.3 to 398.5 minutes.

Why use the 80% figure instead of the 50% figure?

An 80% rate represents four successful attempts out of five in the estimate, while 50% is closer to a coin toss. The report says the Mythos Preview’s 50% horizon was about 17.4 hours, but 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. The benchmark covers a defined set of tasks, and the report says messier work can reduce measured horizons. Companies would need to test their own tasks and account for checking and failed attempts; the supplied evidence does not promise a particular productivity gain.

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

Volaris Reports June 2026 Traffic Results: Load Factor Of 84%

Volaris announced its June 2026 traffic results, achieving an 84% load factor, reflecting strong passenger demand amid ongoing market recovery.

Inside The Engine Room: How Twelve Machines Drive AI

Exploring the 12 core machines behind AI’s functioning, from tokenization to attention, and their impact on AI development and understanding.

Why Talent Density Should Be A Priority In AI Strategy

Exploring how high talent density amplifies AI’s impact, transforming organizational performance and investment priorities in 2026.

Does Google’s Research Prove AI Attacks Can Run Without Humans?

Google says it found a server that automated vulnerability research and exploitation, while broader fully autonomous attacks remain a forecast.