🔍 Read the full analysis: What Does “Time Horizon” Mean For AI And Your Team? 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 latest publicly listed results put an early Claude Mythos Preview at an 80% time horizon of 185.9 minutes, or about three hours, on its benchmark tasks. The measure estimates task length by skilled human work time, not how long the AI takes. Reliability, task messiness and limited public updates mean teams should treat it as a planning reference, not a guarantee.
A November 18 guide from ThorstenMeyerAI.com says teams can use METR’s AI time-horizon measure to estimate which tasks an agent may handle reliably, while warning that the best publicly listed result is about three hours of skilled human work at 80% success on benchmark tasks—not a guarantee for workplace projects. The distinction matters for managers deciding what to delegate, how to break up work and how much human review to budget.
METR, an independent research nonprofit, defines a time horizon as the duration of a task, measured by how long a skilled person takes, at which an AI agent is predicted to succeed at a specified reliability level. The guide focuses on the 80% horizon: an estimated four successes in five attempts. The measure is about task difficulty, not the number of hours the AI itself works; METR says agents are typically several times faster than people on tasks they complete.
According to METR’s Time Horizon 1.1 results file, the early Claude Mythos Preview, dated April 7, 2026, had an 80% horizon of 185.9 minutes. The reported confidence interval is wide, from 97.3 to 398.5 minutes. The guide says the same model’s 50% horizon was 1,044.8 minutes, about 17.4 hours, but METR cautions that estimates above 16 hours are unreliable with its current task suite.
METR’s current benchmark suite contains 228 tasks, according to the guide. Skilled professionals complete the tasks to establish human-time estimates; agents then make six independent attempts per task, with cheating counted as failure. The guide says METR’s public results page is no longer actively updated and was last updated September 8, 2026. The latest model result in the cited file is therefore not a live measure of current AI capabilities.
How Teams Can Use the Measure
For team leads, the time horizon offers a more practical question than whether AI can perform a particular job: how long would a skilled person need, and how dependable is the agent on work of that length? That framing can help with task selection and staffing, but only if managers distinguish benchmark performance from performance in their own code, data and approval processes.
The guide recommends breaking work into hour-scale tasks with clear acceptance checks, assigning testable components first and setting aside time for review. It also advises teams to choose models according to the task and remeasure performance on their own work each quarter. These are recommendations from the guide, not findings from a controlled study of every workplace.
Reliability changes the planning decision. An 80% success estimate leaves one failure in five on comparable benchmark attempts; it does not mean every task will be completed correctly. Teams using agents for consequential work still need a person responsible for checking results, handling failures and deciding when a task is not suitable for automation.
Products Worth Considering

Jackery Fridge Guard 1024Wh Backup Power Station with Mount

Apple AirPods Pro (2nd Generation) Wireless Earbuds with MagSafe Charging Case (Lightning)

Logitech MX Mechanical Mini for Mac Wireless Keyboard, Space Gray
As an affiliate, we earn on qualifying purchases.
How METR Builds Its Estimate
METR’s approach combines human task timing with repeated agent attempts. The guide says the nonprofit uses skilled workers in areas including software engineering, machine learning and cybersecurity, with an average of about five years of relevant experience. Tasks range from seconds to more than eight hours of human work. METR plots success against task length and estimates where the curve falls to a chosen reliability level, such as 80% or 50%.
The public results cited in the guide show the 80% estimate 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, before the Mythos Preview result. Those figures describe performance on METR’s test suite; they do not establish the same pace of progress across all kinds of work. The guide’s longer-term projections are its own extrapolations, not confirmed forecasts.
Benchmark conditions also matter. The guide reports that a METR comparison found messier tasks reduced the horizon by almost a factor of three. A task with clean instructions and a checkable answer may be easier to evaluate than work embedded in an unfamiliar repository, incomplete documentation or changing requirements. That gap helps explain why a polished demonstration and a frustrating everyday result can both occur.
“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 Public Benchmark
The reported 185.9-minute result is an estimate with a wide confidence interval, and METR’s cited public page is no longer actively updated. It is not clear from the guide when a new public results update will appear or how the listed models compare with systems released after April 2026. The guide also describes a trend in which horizons have doubled every three to seven months, but that is a historical fit and not a confirmed timetable for future models.
Results may also shift outside the benchmark. METR tasks are comparatively well specified, while real team work can include unclear requests, dependencies and changing constraints. The source gives no universal conversion from the benchmark to a company’s own workflow. Its reported examples of developer productivity and review accuracy are tied to particular studies and should not be treated as proof that every team will become slower or miss the same share of errors.
Measure Results on Team Work
The immediate next step for teams is to test AI against a sample of their own tasks: record how long skilled staff take, define a pass condition, run the agent and track both completion and review time. The guide recommends repeating that measurement quarterly. This can show whether a task is suitable for automation under a team’s actual conditions, rather than assuming METR’s benchmark transfers directly.
At the research level, readers should watch for an updated METR results file and new evaluations on less tidy tasks. Until such updates are available, the best public figure in the source remains the April 2026 result, and projections beyond it remain uncertain.
Key Questions
Does a three-hour time horizon mean an AI works for three hours?
No. It refers to the amount of skilled human work time associated with a task, not the time the agent spends doing it. METR says agents are typically several times faster than humans on tasks they complete.
What does an 80% time horizon mean?
It is the estimated task length at which an agent succeeds in four out of five attempts on the benchmark. It is a statistical estimate, not a promise that any particular task will succeed.
Which model has the highest result cited in the guide?
The guide identifies an early Claude Mythos Preview, dated April 7, 2026, with an 80% horizon of 185.9 minutes on METR’s benchmark. Its reported confidence interval ranges from 97.3 to 398.5 minutes.
Can teams apply METR’s figure directly to their own backlog?
Not without testing. The guide says benchmark tasks are clean and well specified, while less structured work can reduce measured horizons. Teams should trial agents on representative tasks and include time for human checking and corrections.
Source: ThorstenMeyerAI.com
