🔍 Read the full analysis: What Today’s AI Can Handle, Explained Through Time on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
METR’s public benchmark puts an early Claude Mythos Preview at an 80% time horizon of 185.9 minutes—about three hours of work as measured for skilled humans. The result is not a guarantee: the estimate has a wide confidence interval, and METR warns that its structured tasks may overstate performance on messier work.
A benchmark cited in a November 18 guide from ThorstenMeyerAI.com estimates that an early Claude Mythos Preview can complete tasks requiring a skilled person about three hours of work with 80% reliability. The measure, developed by research nonprofit METR, is intended to help teams judge which tasks may be suitable for AI agents; it does not mean the system works for three hours or succeeds on every task.
METR defines a model’s time horizon as the duration of tasks, measured by human expert completion time, at which an AI agent is predicted to succeed at a specified rate. In the cited Time Horizon 1.1 results file, Mythos Preview’s 80% estimate is 185.9 minutes. The guide says the estimate’s confidence interval runs from 97.3 to 398.5 minutes, roughly 1.6 to 6.6 hours.
The benchmark uses skilled professionals to time tasks, then gives AI agents repeated attempts. METR’s current task suite contains 228 tasks; the guide says METR runs each task six times and scores cheating as failure. Researchers fit a success curve across task lengths and report where it crosses the selected reliability threshold. The 80% figure means roughly four successful attempts out of five at the estimated task length, not a guarantee for a particular assignment.
The guide reports that the same model’s estimated horizon falls to 3.1 hours at 80% success, compared with 17.4 hours at 50%. It cautions that the longer estimate is uncertain: METR says measurements above 16 hours are unreliable with its current task suite. The results file lists earlier milestones including 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.
How Teams Can Use the Estimate
The benchmark frames AI capability in terms of task duration and estimated reliability: how long a skilled person needs for a task, and how often a model completes tasks of that length. This may help teams evaluate which work to assign to agents, set review requirements and determine the size of individual assignments.
The guide recommends dividing work into short, clearly specified pieces, starting with tasks that have testable acceptance criteria, and budgeting time for human review. It also advises teams to test models on their own work and repeat those measurements regularly. These are recommendations in the guide, not findings that every organization will see the same performance.
At the 80% threshold, the estimate corresponds to a failure rate of about one in five attempts at the measured task length. Teams using AI for software or other consequential work need to determine who will check the output and how errors will be handled.
Products Worth Considering

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

DJI Mini 4K Camera Drone Combo with 2 Batteries

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Length
METR’s approach is built around human working time, not the time an AI spends generating an answer. The guide says its task authors include professionals in software engineering, machine learning and cybersecurity, with an average of about five years of relevant experience. Their completion times provide the scale for tasks ranging from seconds to more than eight hours.
METR then tests agents on those tasks and examines their transcripts. Success rates generally decline as task duration rises; the reported horizon is the point on the fitted curve corresponding to a chosen success rate. The guide focuses on 80% because it treats that level as more useful for planning than a 50% success rate, while stressing that neither threshold guarantees success on an individual task.
The source also says METR’s public results page was marked no longer actively updated, with a last update of September 8, 2026. Its cited benchmark file contains the April 2026 Mythos Preview result. Separately, METR’s February-to-March 2026 Frontier Risk Report put the labs’ internal frontier at at least three but under four hours at 80%, according to the guide.
“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
Limits of the Three-Hour Estimate
The 185.9-minute figure is an estimate from a defined benchmark, not a measurement of all workplace tasks. Its reported confidence interval is wide, and the source says messier tasks reduced the horizon by almost a factor of three in one METR comparison. The guide does not provide enough detail in the supplied material to assess how broadly that comparison applies.
It is also unclear how the listed result translates to a particular company’s code, review processes, or other work. The guide says the public benchmark page was no longer actively updated and that nothing after April 2026 had been measured publicly in the cited record. Its forecast that an eight-hour workday horizon could arrive between September 2026 and November 2027 is the author’s own fit of past trends, not a confirmed benchmark result or a guarantee.
The supplied source material begins to describe findings about developer speed and error review but cuts off before presenting the full study details. Those claims cannot be fully evaluated from the material provided, so they should not be treated here as settled evidence about every team’s productivity.
Testing Models on Real Work
The guide recommends that organizations compare models against their own tasks, use clear success criteria and record how much human checking is needed. It recommends repeating those evaluations each quarter as tools and workflows change. METR’s public benchmark status, as described in the guide, means readers should check for newer results before treating the April 2026 record as current.
The three-hour figure describes estimated performance on METR’s task suite at an 80% success threshold. Whether that performance carries over to a team’s less structured work remains an open question.
Key Questions
What does an AI time horizon measure?
It estimates the length of a task, based on how long skilled humans take, at which an AI agent reaches a specified success rate. It does not measure the time the AI spends completing the task.
What is the reported result for Claude Mythos Preview?
The cited METR results file lists an 80% time horizon of 185.9 minutes for an early Claude Mythos Preview dated April 7, 2026. The guide reports a confidence interval of 97.3 to 398.5 minutes.
Does a three-hour horizon mean the AI will finish every three-hour task?
No. It is a fitted estimate for a task length and a chosen reliability level. At the 80% threshold, the estimate corresponds to success on about four out of five attempts, not every attempt.
Why does the task’s setting matter?
METR’s tasks are structured and well specified, according to the guide. It reports that messier tasks reduced the horizon by almost a factor of three in one comparison, so results may not transfer directly to loosely defined workplace assignments.
How should teams apply the benchmark?
The guide recommends breaking work into short tasks with clear acceptance checks, reserving time for review, and testing models on a team’s own work. The benchmark is a reference point, not a substitute for local evaluation.
Source: ThorstenMeyerAI.com
