The AI Time Horizon: Measuring Progress One Hour At A Time
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: The AI Time Horizon: Measuring Progress One Hour At A Time on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day 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 records put an early Claude Mythos Preview at a 185.9-minute task horizon with an 80% success rate, based on results dated April 7, 2026. That is a measure of task difficulty in skilled human work time, not how long the AI takes to work; the benchmark is limited, and results after April 2026 are not publicly measured in the cited material.

A November 18 guide from ThorstenMeyerAI.com says the best result in the cited public METR benchmark is an 80% task horizon of 185.9 minutes for an early Claude Mythos Preview, meaning the model is predicted to complete tasks that take a skilled human about three hours four times in five. The measure is intended to help teams judge which work AI can handle, but it is not a guarantee of success on any individual task or a claim that the model works for three hours.

The result is listed in METR’s Time Horizon 1.1 results file with a date of April 7, 2026. The guide says the record 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, before reaching 185.9 minutes for the early Mythos Preview. These are benchmark measurements, not direct evaluations of every kind of workplace task.

METR defines a time horizon using the amount of skilled human working time a task requires and the probability that an AI agent completes it. Researchers time professionals on tasks, have AI agents attempt those tasks in multiple runs, and fit a success curve across task lengths. The guide says METR’s current Time Horizon 1.1 suite contains 228 tasks; agents get six independent runs per task, and attempts judged to involve cheating count as failures.

At the same model result, the estimated horizon changes with the selected reliability threshold. The guide reports a 50% horizon of 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 the current task suite. The reported 80% estimate also has a wide confidence interval, from 97.3 to 398.5 minutes.

At a glance
reportWhen: Guide published November 18, 2026; cite…
The developmentA November 18, 2026 guide from ThorstenMeyerAI.com explains METR’s task-time benchmark and reports a 185.9-minute, 80%-reliability result for Claude Mythos Preview.

Using the Horizon to Plan Work

The measure provides a way to compare task duration, measured in skilled human work time, with a model’s estimated success rate. For tasks with clear specifications and verifiable outputs, a horizon estimate may inform decisions about delegation, work breakdown and review requirements.

The benchmark does not by itself predict productivity on real projects. The guide says METR’s tasks are cleaner and more clearly specified than many workplace assignments; in one METR comparison, messier tasks reduced the horizon by almost a factor of three. Teams also need to account for failed attempts and verification. The guide cites a 2025 METR trial in which experienced developers were 19% slower with AI while believing they were 20% faster. That finding applies to the cited trial and does not establish how other teams or tools will perform.

The guide recommends dividing work into hour-scale pieces with explicit acceptance checks, starting with testable tasks, and measuring performance on a team’s own work. Local results are needed to assess whether the benchmark applies to a particular organization’s delivery estimates.

Products Worth Considering

Therabody Theragun Sense 2nd Generation Percussion Massage Gun, Black

Therabody Theragun Sense 2nd Generation Percussion Massage Gun, Black

As an affiliate, we earn on qualifying purchases.

How METR Measures Task Length

METR is an independent research nonprofit that evaluates AI systems on tasks including software engineering, machine learning and cybersecurity. The guide says the human participants are skilled professionals with an average of about five years of relevant experience. Their completion times provide the human-time scale against which agents’ performance is compared.

The reported horizon is the point where a fitted success curve reaches a chosen reliability level. At the 80% threshold, four out of five attempts are expected to succeed across the measured task range; at 50%, success is less reliable. The figure is therefore about task difficulty and estimated success probability, not the duration of an AI work session. METR says agents can complete tasks they solve several times faster than humans.

The figures also have a publication limit. The guide reports that METR’s public results page says it is no longer actively updated and was last updated September 8, 2026. It cites a METR Frontier Risk Report for February to March 2026 estimating the labs’ internal frontier at at least three but less than four hours at 80% reliability. The guide’s cited public record ends with the April 2026 Mythos Preview result.

“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 guide

Benchmark Limits and Missing Updates

The cited 185.9-minute result has a wide confidence interval, and the guide says the public METR page is no longer being actively updated. It reports that no public measurement after April 2026 was available in the material it reviewed; the current capabilities of newer or internally tested models therefore cannot be established from this record.

It is also unclear how well the benchmark predicts performance on a particular company’s work. METR’s task suite cannot represent every codebase, ambiguous assignment, changing requirement or review process. The guide cites a separate study in which reviewers caught only 31% of errors that required knowledge of a task’s rules to identify, underscoring that a nominal pass rate does not remove the need for capable oversight.

The cited results do not establish that AI can complete a full workday reliably. The guide gives a forecast range for when an eight-hour task horizon at 80% might be reached, but says the estimate depends on which years of progress are used. That forecast is not a measured result, and future performance remains uncertain.

Track Results on Real Work

The guide recommends that teams begin with bounded, testable tasks, define acceptance criteria in advance, and include review time in project estimates. It also advises choosing models task by task rather than assuming one benchmark result applies across different work. These steps would let teams compare agent performance with their own task durations and quality requirements.

For researchers and readers tracking broader capability trends, the next useful evidence would be updated public benchmark results, clearer reporting on task-suite limits, and repeated tests on less structured work. Until then, the three-hour figure should be read as a dated benchmark estimate with a defined reliability threshold, not a current, universal measure of what AI can do.

Key Questions

What does a three-hour AI time horizon mean?

It means the model is estimated to succeed with a specified reliability on tasks that take a skilled human about three hours. It does not mean the AI works for three hours or completes every such task.

Which model had the highest result in the cited METR data?

The guide identifies an early Claude Mythos Preview, with a 185.9-minute horizon at 80% reliability in a result dated April 7, 2026. The estimate’s confidence interval is 97.3 to 398.5 minutes.

Why is the 50% time horizon much longer?

A 50% threshold allows a lower estimated success rate than 80%, so it reaches tasks of greater human duration. The guide reports 1,044.8 minutes at 50% for the same model, but METR warns that measurements above 16 hours are unreliable with its current task suite.

Can companies use the benchmark to estimate productivity gains?

It can inform task selection, but it does not establish productivity gains for a particular organization. Task clarity, error checking, failed attempts and local workflows all affect results, so teams should measure performance on their own work.

Are there newer public results after April 2026?

The guide says the public METR results page was no longer actively updated and that the cited public record ended with the April 2026 result. It does not establish whether newer internal measurements exist or what those results show.

Source: ThorstenMeyerAI.com

You May Also Like

The Machine Economy — Capital-Heavy, Human-Light, Trading With Itself

Analysis of the emerging machine economy where AI-driven firms trade among themselves, operate with minimal human input, and reshape the global economic landscape.

How Suzanne Ciani’s Buchla Cookbook Shapes The Future Of Live Event Signals

A new workflow inspired by Suzanne Ciani’s Buchla Cookbook is being tested to help promoters catch live show developments early and make informed decisions.

The Safety Card, Played From Every Side: David Sacks, Anthropic, and the Fable Standoff

White House officials and Anthropic dispute the nature of a cybersecurity jailbreak in their models, raising questions about AI safety and transparency.

Transform Your Workflow In 2026 With These 15 AI Tools

Discover 15 top AI tools for workflow automation in 2026, designed to boost productivity, streamline tasks, and seamlessly integrate with existing systems.