🔍 Read the full analysis: Measuring AI Progress: How Long Can It Handle A Task? on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s public benchmark results put an early Claude Mythos Preview at an 80% task horizon of 185.9 minutes, or about three hours, for the tasks in its test suite. The measure estimates how long a skilled human would need for tasks an AI agent completes at a specified success rate; it does not mean the AI works at human speed or succeeds reliably on every three-hour job. The estimate has a wide confidence interval, and METR’s public results page is no longer actively updated.
METR’s public task benchmark puts an early Anthropic Claude Mythos Preview at an 80% time horizon of 185.9 minutes—about three hours—according to results reported by ThorstenMeyerAI on November 18, 2026. The figure estimates the length of tasks, measured by skilled human working time, that an AI agent completes four times out of five on METR’s test suite; it is not a guarantee for everyday workplace projects.
METR, an independent research nonprofit, calculates its time horizon by timing skilled people on tasks, asking AI agents to attempt those tasks, and estimating how task length relates to success. Its Time Horizon 1.1 suite contains 228 tasks across areas including software engineering, machine learning and cybersecurity. The source says METR uses six independent runs per task and treats detected cheating as failure.
The reported record is for an early version of Claude Mythos Preview dated April 7, 2026 in METR’s results file. Its 80% estimate is 185.9 minutes, with a confidence interval from 97.3 to 398.5 minutes. The same model’s estimated horizon at 50% success is 1,044.8 minutes, or about 17.4 hours, though METR cautions that measurements above 16 hours are unreliable with its current task suite.
The figures describe benchmark performance, not a broad test of office work. The source reports that METR’s tasks are relatively clean and well specified and cites a comparison in which messier tasks reduced the horizon by almost a factor of three. It also says the public results page was last updated September 8, 2026, and is no longer actively updated. The cited record therefore should not be read as a current, continuously maintained ranking.
Using Task Length to Plan AI Work
The horizon offers teams a more practical question than whether AI can perform a task at all: how long would the task take a skilled person, and how often does the agent finish comparable work? A three-hour estimate can help managers decide which bounded tasks to trial, but the success rate matters as much as the duration. At the 80% threshold, roughly one in five attempts may fail under the benchmark’s conditions.
That failure rate has operational costs. A person may need to inspect results, diagnose mistakes and redo work. The source argues that teams should split assignments into pieces with clear acceptance checks, begin with work that can be tested, and budget time for review. It also recommends measuring performance on a team’s own tasks rather than assuming benchmark results transfer directly.
The measure is not a forecast that a model can complete an entire workday independently. A longer horizon at 50% reliability is a weaker basis for routine delegation than a shorter horizon at 80%, and neither percentage means every task of that length has the same chance of success. For businesses, the useful implication is to test task-specific workflows and track both completion and oversight costs.
Products Worth Considering

Samsung 27-Inch Odyssey G5 QHD Gaming Monitor (G50F)

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, White Smoke

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen
As an affiliate, we earn on qualifying purchases.
How METR Calculates the Horizon
METR defines the time horizon as the duration of tasks, measured by human expert completion time, at which an AI agent is predicted to succeed with a chosen level of reliability. The human-time measure is about task difficulty, not how long the AI itself spends working. The source says agents can complete successful tasks several times faster than people.
Results vary with the selected reliability threshold. The 80% horizon identifies the task length at which the fitted success curve crosses four successes in five; the 50% horizon marks a coin-flip level. The source uses 80% as a planning reference, while emphasizing that it still allows failures. METR’s confidence interval for the reported record is broad, spanning roughly 1.6 to 6.6 hours.
The reported series rose 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 cited Mythos Preview result. These are benchmark estimates at the 80% threshold, not direct measurements of productivity across all uses. ThorstenMeyerAI also notes that the apparent pace of growth depends on which years are used to fit the trend; forecasts based on that history remain projections, not measured outcomes.
“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 report
Benchmark Limits and Missing Updates
The estimate applies to METR’s particular task suite, not to all jobs or software projects. The source says messier tasks can sharply reduce measured horizons, but does not provide enough detail here to determine how the reported comparison would translate to a specific company’s workflow. The confidence interval for the 185.9-minute estimate is also wide, and the benchmark’s 50% estimate exceeds the range METR says it can measure reliably.
There is no public update after the listed April 7, 2026 result in the material cited by the report. METR’s page was last updated on September 8, 2026, and is no longer actively updated. It is consequently unclear from these sources whether later models have longer horizons, how results compare across changing benchmark versions, or how much review time people need on the measured tasks.
The source also cites separate research on human use of AI: experienced developers in a 2025 trial were 19% slower with AI while believing they were 20% faster, and reviewers in another study caught 31% of errors requiring knowledge of task rules. Those findings are presented as cautions, but the supplied material does not give enough study detail to generalize them to every team or workflow.
Testing Results on Real Team Tasks
The immediate next step for teams is local measurement: select representative tasks, define what counts as a correct result, and record both agent success and the time spent checking or fixing work. Repeat the test as models and workflows change. The source recommends reviewing performance quarterly, but does not identify a scheduled next release of METR’s public benchmark.
For the field, a useful next milestone would be updated, comparable measurements that make the model version, task suite, reliability threshold and uncertainty clear. Until then, the reported three-hour figure is best treated as a dated benchmark result and a starting point for controlled trials—not a promise that an AI agent can take over a three-hour assignment without supervision.
Key Questions
What does an AI time horizon measure?
It estimates the length of a task, measured by skilled human working time, at which an AI agent reaches a specified success rate on a benchmark. It does not measure how long the AI takes to do the work.
What is the reported 80% result?
The cited METR results put an early Claude Mythos Preview at 185.9 minutes, or about three hours, on the Time Horizon 1.1 benchmark. The estimate is dated April 7, 2026 and has a confidence interval of 97.3 to 398.5 minutes.
Does a three-hour horizon mean the model will finish any three-hour task?
No. It is an estimated success threshold for tasks in METR’s benchmark, not a guarantee for an individual task or a different workplace setting. At an 80% threshold, some attempts still fail.
Why is the 50% horizon much longer?
The same model’s reported horizon is about 17.4 hours at 50% success, compared with about 3.1 hours at 80%. A lower success threshold permits longer tasks, but METR warns that measurements above 16 hours are unreliable with its current task suite.
Is this the latest AI benchmark result?
The supplied report says METR’s public results page was last updated September 8, 2026, and is no longer actively updated. The cited model result is from April 2026, so these sources do not establish which model has the longest horizon now.
Source: ThorstenMeyerAI.com
