🔍 Read the full analysis: AI Today, By The Hour: A Guide To Its Time Horizon on ThorstenMeyerAI.com
Get audio and creator gear 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 lists an 80% time horizon of 185.9 minutes for an early Claude Mythos Preview released in April 2026. That means success on tasks taking skilled people about three hours, not that the AI works for three hours or can reliably complete any three-hour job. The result is uncertain, based on a specific task suite, and the public benchmark has not been actively updated since September 2026.
A guide published by ThorstenMeyerAI.com on November 18, 2026, says the best publicly measured AI agent completed tasks requiring about three hours of skilled human work with an estimated 80% success rate. The figure comes from METR, an independent research nonprofit, and refers to an early Claude Mythos Preview released in April 2026; it is a benchmark result, not a guarantee that the system can handle any three-hour assignment.
METR defines a model’s time horizon as the human expert time needed to complete tasks at a specified level of AI reliability. Its Time Horizon 1.1 benchmark gives the Claude Mythos Preview an 80% horizon of 185.9 minutes, or just over three hours. The source article says METR’s results file dates the model’s result to April 7, 2026. At this threshold, four successes in five attempts are estimated for tasks around that duration.
The benchmark’s earlier records show the measured horizon 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. Those figures describe results on METR’s task suite. They do not establish that the same performance applies to every workplace, software project or task type.
The guide reports a wide confidence interval for the Mythos estimate: 97.3 to 398.5 minutes, or roughly 1.6 to 6.6 hours. It also says METR’s public results page was marked no longer actively updated, with its last update on September 8, 2026. The guide cites a separate METR report estimating the labs’ internal frontier at “≥3h but <4h” at 80% for February to March 2026.
Turning Benchmarks Into Work Plans
The time-horizon measure offers managers a way to assess how long tasks take skilled people and how often an AI completes them. Those results can inform decisions about where to test an agent, how to divide work and how much review time to allocate.
An 80% success estimate indicates that failures remain possible. It does not mean that exactly four of every five tasks assigned by a particular team will succeed, or that the benchmark’s rate will apply to that team’s work. The guide recommends dividing assignments into short, clearly specified pieces, beginning with tasks that can be checked through tests, and allowing time for human review.
Reliability changes the estimated horizon. For the same Mythos Preview, the guide cites a 50% time horizon of 1,044.8 minutes, about 17.4 hours, compared with 185.9 minutes at 80%. METR cautions that measurements above 16 hours are unreliable with its current task suite, so the longer estimate has that limitation.
Products Worth Considering

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity

Anker SOLIX C300 Portable Power Station, 288Wh
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Length
METR’s method starts by having skilled professionals complete tasks and recording their working time. The source article says the contractors work in software engineering, machine learning and cybersecurity and average about five years of relevant experience. The current Time Horizon 1.1 suite contains 228 tasks, spanning seconds to more than eight hours of human work.
METR then runs an AI agent on each task multiple times. The guide says the benchmark uses six independent runs per task and treats cheating identified in transcripts as failure. Researchers plot success rates against human task duration and estimate where the curve crosses a chosen reliability level, such as 80% or 50%.
The resulting number measures task difficulty in human time, not how long the AI itself takes. The guide says agents are typically several times faster than people on tasks they complete. That distinction matters: a three-hour horizon does not mean the system spends three hours working, nor that it can sustain a human workday of independent, dependable output.
“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 benchmark’s confidence interval is broad, and its results depend on a defined set of tasks and testing conditions. The source guide says that messier tasks cut the horizon by almost a factor of three in one METR comparison. It does not provide enough detail here to establish how that adjustment applies across different industries or individual teams.
METR’s public record is not actively updated, according to the guide, so the cited April 2026 result does not show how models released or tested later performed. The guide also mentions a 2025 METR trial in which experienced developers were 19% slower with AI while believing they were 20% faster, and a separate study in which reviewers caught 31% of errors requiring task-specific rules. Those findings are reported by the source article; their study details are not included in the supplied material.
It is also unclear how closely a company’s tasks match METR’s benchmark, what success threshold it should require for its own work, or how much checking a particular AI-generated result will need. The guide’s projections of when an eight-hour task horizon might be reached depend on trend fits, not measurements after April 2026.
Testing Against Team Backlogs
The guide recommends teams start with their own work rather than use the public benchmark as a direct forecast. They can split assignments into hour-scale tasks with clear acceptance checks, try AI first on parts that can be tested, and record both completion rates and the time needed to review and correct results.
ThorstenMeyerAI.com also recommends choosing models by task and re-measuring performance on internal work every quarter. Any forecast that the benchmark will reach a full eight-hour workday should be treated as a projection: the guide’s fits place that point between September 2026 and November 2027, depending on the years included, while noting that no public measurements after April 2026 are available in the cited record.
Key Questions
What does an AI time horizon measure?
It estimates the length of a task, in skilled human working time, at which an AI agent reaches a stated success rate. It does not measure how long the AI takes to perform the task.
What is the latest public result cited in the guide?
The guide cites an 80% time horizon of 185.9 minutes for an early Claude Mythos Preview released in April 2026. It reports a confidence interval of 97.3 to 398.5 minutes.
Does a three-hour horizon mean the AI can do any three-hour job?
No. The result is based on METR’s task suite and an estimated success rate. The guide says messier tasks performed worse in one comparison, and a team’s results may differ.
Why is the 50% horizon much longer?
A lower success threshold allows longer tasks. The guide cites about 17.4 hours at 50% for the same model, versus about 3.1 hours at 80%, while noting METR considers measurements above 16 hours unreliable with its current suite.
What should teams do with the benchmark?
Use it as a starting point, not a guarantee: divide work into testable pieces, track performance on the team’s own tasks, and include time for human review and correction.
Source: ThorstenMeyerAI.com
