🔍 Read the full analysis: AI Capability In Hours: What Teams Need To Know 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
METR’s latest public results put the 80% task horizon of an early Claude Mythos Preview at 185.9 minutes, meaning it succeeded on about four out of five benchmark attempts at tasks taking skilled humans roughly that long. The result is not a promise of three hours of autonomous work: METR’s tasks are well specified, reliability falls as tasks get longer, and the public benchmark has not been updated with results after April 2026.
A report published by ThorstenMeyerAI.com on November 18, 2026, uses METR’s task-duration benchmark to advise teams on where AI agents may fit into work planning. METR’s latest cited public result gives an early Claude Mythos Preview an 80% time horizon of 185.9 minutes: on the benchmark, the model succeeded about four times in five on tasks that skilled people take roughly three hours to complete. The figure is a measure of task difficulty and success rate, not a guarantee that an AI can work independently for three hours.
METR, an independent research nonprofit, estimates how long a skilled human would need for a task, then measures how often AI agents complete tasks of different lengths. Its current cited suite, Time Horizon 1.1, contains 228 tasks. METR uses skilled professionals, averaging about five years of relevant experience, to establish human completion times. It then runs AI systems on tasks multiple times and evaluates their work; the reported setup includes six independent runs per task, with cheating counted as failure.
The report says the early Claude Mythos Preview result is dated April 7, 2026, in METR’s results file. Its estimated 80% horizon is 185.9 minutes, with a wide confidence interval of 97.3 to 398.5 minutes. Earlier results cited in the same file include 0.89 minutes for GPT-4 in March 2023, 29.98 minutes for o3 in April 2025 and 66 minutes for GPT-5.2 in December 2025. These figures describe performance on METR’s benchmark, not a universal ranking across every kind of work.
The report recommends that teams break work into tasks measured in hours, define clear acceptance checks, start with parts that can be tested, and account for human review. It also advises teams to compare models on their own work and repeat those checks regularly. These are the report’s recommendations, not findings that every organization will see the same results.
How Teams Can Use the Horizon
The measure offers managers a more practical question than whether AI can perform a task at all: how long is the task, and how reliably does the system complete work of that length? That can inform how teams divide backlogs, choose which tasks to delegate, and set review capacity. A short, testable task with a clear pass-or-fail check is easier to evaluate than a broad request with ambiguous requirements.
But the benchmark does not remove the need for oversight. At an 80% success rate, roughly one in five attempts fails under the measured conditions. A team using the horizon for planning still needs to account for failed work, checking time, and the consequences of an error. The report also cites separate research in which experienced developers using AI were 19% slower while believing they were 20% faster, and a review study in which reviewers caught 31% of errors requiring knowledge of task-specific rules. Those findings are attributed in the source report; their applicability will depend on the tasks and teams involved.
Products Worth Considering

Canon SELPHY CP1500 Compact Photo Printer Kit with 2x KP-108IN Ink and Paper Sets

Canon SELPHY QX20 Compact Photo Printer – Dark Gray

FLEXISPOT 3-in-1 Convertible Chair and Table for Four
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Length
METR defines a time horizon as the task duration, measured by human expert completion time, at which an AI agent is predicted to succeed with a specified reliability. The 80% horizon is the estimated task length where success falls to four out of five attempts; the 50% horizon marks a lower reliability level. The number is not the time the AI takes to finish. According to METR, agents typically complete tasks they succeed at several times faster than people.
Reliability changes the headline number. For the same Mythos Preview, the report gives a 50% horizon of 1,044.8 minutes, or about 17.4 hours, compared with 185.9 minutes at 80%. METR cautions that estimates above 16 hours are unreliable with its current task suite. The report also says messier tasks reduced the measured horizon by almost a factor of three in one METR comparison, a reminder that clean benchmark tasks may not reflect the interruptions and ambiguity of workplace projects.
The report describes the trend line as having doubled every three to seven months, if it continues. Its own projections put an eight-hour workday at 80% somewhere between September 2026 and November 2027, depending on which period of data is used. That is a forecast based on historical measurements, not an observed result. The report says nothing after April 2026 had been publicly measured in the cited record.
Limits of the Public Benchmark
The cited record is not a live account of current frontier capability. METR’s public results page was last updated on September 8, 2026, and the newest model entry cited in the report is from April 7. The report also references METR’s February-to-March 2026 Frontier Risk Report, which placed the labs’ internal frontier at “≥3h but <4h” at 80%; that is a separate estimate, not a later public benchmark result.
It remains unclear how the reported horizon translates across different industries, internal codebases, and tasks with less precise instructions. The confidence interval around the Mythos Preview estimate is broad, and METR says its task suite cannot reliably measure horizons above 16 hours. The supplied material does not provide results from an independent replication of the specific 185.9-minute figure or a newer public measurement that would confirm the forecasted trend continued.
What Teams Should Measure Next
For organizations, the next useful step is to test AI on representative internal tasks rather than treat a public horizon as a deployment threshold. Teams can record the human time required, define what counts as a correct result, measure repeated attempts, and include the time spent reviewing and repairing outputs. The report recommends repeating these measurements every quarter and choosing models according to task needs.
For the broader capability picture, a key milestone would be renewed public benchmark data from METR or comparable evaluations. Until that appears, the April 2026 result and projections based on earlier trends should be treated as the latest cited measurement, not confirmation of current performance. Whether AI can complete longer workplace tasks reliably—and at what review cost—remains an open question.
Key Questions
What does an 80% AI time horizon mean?
It is the estimated length of a task, based on skilled human completion time, at which an AI agent succeeds on about four out of five attempts in METR’s benchmark. It does not mean the AI spends that amount of time working.
Which model has the highest result cited in the report?
The report identifies an early Claude Mythos Preview, dated April 7, 2026 in METR’s results file, with an 80% horizon of 185.9 minutes. Its confidence interval is wide, and the cited public record has not been updated with later measurements.
Can teams expect AI to complete a three-hour task reliably?
No benchmark result guarantees success on a particular workplace task. The reported figure is an estimate across METR benchmark tasks, and an 80% rate still implies failures. Teams need task-specific tests and human review.
Why is the 50% horizon much longer?
A lower success threshold allows longer tasks into the estimate. For the cited Mythos Preview, the report gives a 50% horizon of about 17.4 hours, compared with about 3.1 hours at 80%. METR says measurements above 16 hours are unreliable with its current suite.
What should a team do with the benchmark?
Use it as context, then test representative work internally. Break tasks into manageable pieces, set explicit acceptance checks, measure failure and review rates, and reassess performance periodically rather than assuming benchmark results transfer directly.
Source: ThorstenMeyerAI.com
