🔍 Read the full analysis: How To Use The Time Horizon To Assess AI For Your Team on ThorstenMeyerAI.com
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
TL;DR
An analysis published Nov. 18 explains how METR’s “time horizon” measures the length of human tasks AI agents can complete at a specified success rate. The source reports a record of 185.9 minutes at 80% for an early Claude Mythos Preview, while warning that benchmark conditions, reliability and human review affect how teams should use the figure.
METR, an independent research nonprofit, defines a time horizon as the duration of a task, measured by how long a skilled human expert takes, at which an AI agent is predicted to succeed at a given reliability level. The source article uses the 80% success threshold as its planning measure: four successful attempts out of five, rather than a guarantee that any individual task will succeed.
According to the article’s account of METR’s Time Horizon 1.1 results file, the record cited is an early Claude Mythos Preview, dated April 7, 2026, with an 80% horizon of 185.9 minutes. The same file lists 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 measurements, not evidence that a model can complete every workplace task of equal duration.
The article says the 185.9-minute estimate has a wide confidence interval, from 97.3 to 398.5 minutes. It also reports a 50% horizon of 1,044.8 minutes, or about 17.4 hours, for the same model. METR cautions that measurements above 16 hours are unreliable with its current task suite, according to the source.
Turning Model Scores Into Work Plans
The time-horizon framing gives managers a more practical question than whether AI can perform a task in principle: how long does the task take an experienced person, and how often does the agent finish it? That can help teams decide which work to trial, where to place human checks and how much time to allocate for review.
The benchmark does not establish that a three-hour human task can be safely delegated in every workplace. Its 80% rate implies failures remain possible, while the source says METR’s tasks are cleanly specified and that messier tasks in a comparison reduced the horizon by nearly a factor of three. Teams therefore need to account for task ambiguity, error costs and verification effort, not just task duration.
The source also cites a METR 2025 trial in which experienced developers working with AI were 19% slower, although they believed they were 20% faster. It references a separate study where reviewers caught 31% of errors requiring knowledge of task rules. Those findings, as presented by the source, caution against treating perceived speed or model output as proof of productivity.
Products Worth Considering

Apple Mac mini Desktop Computer with M6 Chip (2026)

Apple Mac mini with M2 chip, 16GB unified memory, 256GB SSD, Silver

Logitech MX Brio Ultra HD 4K Webcam
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Length
The measurement process described in the article starts with skilled professionals completing tasks. METR records human working time on tasks in areas including software engineering, machine learning and cybersecurity; participants average about five years of relevant experience. The current Time Horizon 1.1 suite cited in the article contains 228 tasks.
METR then runs an AI agent six times on each task and checks transcripts for cheating, which counts as failure. It plots success rates against human task duration and fits a curve. The point where that curve reaches a chosen success rate produces the horizon: 80% for the article’s planning figure, or 50% for a less reliable threshold.
The horizon measures task difficulty in human-time units, not how long the AI itself takes to work. The source says agents are typically several times faster than humans on tasks they complete. METR’s public results page, however, was marked as no longer actively updated, with a last update of Sept. 8, 2026, according to the article.
Limits Behind The Three-Hour Figure
The latest result cited is dated April 7, 2026, and the article says no public measurements after that date were available. METR’s public results page was no longer being actively updated, according to the source, so the cited record does not establish the state of models released later.
The estimate also has a broad confidence interval, and the underlying tasks may not reflect the ambiguity, dependencies or shifting requirements of a team’s live work. The article reports a near-threefold reduction on messier tasks in one METR comparison, but it does not provide enough detail here to show how that result applies across different industries or task types. How well the benchmark predicts performance on a particular team’s work remains uncertain.
Nor does an 80% horizon say that every task at that duration will succeed four times in five for every user or setting. It is a fitted benchmark estimate across task attempts. Review demands, access to tools and data, and the cost of a mistaken result can all change whether delegation makes sense.
Testing The Measure On Team Work
The article recommends starting with work that can be divided into hours-long pieces with clear acceptance checks, then assigning AI the parts that are easiest to test. Teams should budget time for human review, select models according to the task and remeasure performance on their own work every quarter.
That approach would give managers local evidence alongside the public benchmark: which task types succeed, where errors occur, and whether checking offsets any speed gains. The source does not name a next public METR update or a date for new benchmark results. Until further measurements are available, the April 2026 result is a historical benchmark, not a forecast or guarantee of what a team’s tools can do next.
Key Questions
What does an AI time horizon measure?
It estimates the length of a task, measured by skilled human completion time, at which an AI agent reaches a specified success rate. It does not measure how long the AI takes to perform the task.
What is the latest 80% figure cited in the article?
The source reports 185.9 minutes for an early Claude Mythos Preview, based on a METR results file entry dated April 7, 2026. The reported confidence interval is 97.3 to 398.5 minutes.
Does a three-hour horizon mean AI can reliably do any three-hour task?
No. It is a benchmark estimate at a particular success threshold, based on METR’s task suite. The source warns that messier tasks can have a shorter horizon and that an 80% rate still leaves room for failure.
How should a team apply the measure?
The article suggests breaking work into clear, testable pieces, starting with tasks that are easy to check, budgeting for human review and measuring results on the team’s own work. It recommends repeating those measurements quarterly.
Source: ThorstenMeyerAI.com
