🔍 Read the full analysis: How Long Can AI Work Independently? A Guide For Teams on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s public results put the best measured model’s 80% time horizon at 185.9 minutes, based on tasks timed for skilled human workers. The benchmark is not a promise that AI can work unsupervised for three hours: reliability falls on longer tasks, and messier work can shorten the measured horizon. The public record is dated April 2026, so it does not show how models released afterward perform.
METR’s latest public results put the best measured AI model’s 80% task horizon at 185.9 minutes, or about three hours, according to a November 18 guide from ThorstenMeyerAI.com. The measure describes the length of tasks, in skilled human working time, that an AI agent completes at a specified success rate; it does not mean the system can work independently for that many hours without supervision.
The guide draws on METR, an independent research nonprofit, which measures task length using the time skilled people take to complete work. Its Time Horizon 1.1 suite contains 228 tasks, according to the guide. METR has professionals attempt the tasks, runs AI agents multiple times, and estimates how success changes as tasks become longer. The 80% horizon is where the fitted success curve reaches four successful attempts out of five.
The cited record is an early version of Anthropic’s Claude Mythos Preview, dated April 7, 2026 in METR’s results file. Its 80% horizon is listed as 185.9 minutes, with a wide confidence interval of 97.3 to 398.5 minutes. The guide compares that result with 0.89 minutes for GPT-4 in March 2023 and 29.98 minutes for o3 in April 2025. These are benchmark measurements, not evidence that every team’s software or business tasks will have the same success rate.
The guide says the same model’s 50% horizon is 1,044.8 minutes, about 17.4 hours, but cautions that this is a less dependable planning figure. METR warns that measurements above 16 hours are unreliable with its current task suite. The guide also reports that METR’s public results page is no longer actively updated and was last updated September 8, 2026, leaving newer model performance unreported in this dataset.
Planning Work Around AI Reliability
For team leads, the benchmark offers a more practical question than whether AI can perform a task at all: how long would a skilled person need, and how often does the agent finish work of that length? That can help teams break projects into pieces, assign testable tasks, and estimate how much review capacity they need.
But an 80% result still implies one failure in five attempts around the benchmark’s estimated crossing point. A team that treats the number as a guarantee risks allowing errors to move into later work. The guide recommends starting with tasks that have clear acceptance checks, budgeting for human review, selecting models by task, and measuring outcomes on the team’s own work. Those are practical proposals from the guide, not findings that METR has tested for every workplace.
The difference between a clean benchmark task and a live backlog also matters. A task can be well specified in a test suite but depend in a company on undocumented code, shifting requirements, or decisions that require product knowledge. Those conditions affect whether the benchmark translates into dependable work, and the guide says messier tasks can reduce the horizon substantially.
Products Worth Considering

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite

Apple Watch Series 12 GPS Smartwatch, 42 mm

Samsung 32 Odyssey G5 QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Length
METR defines its 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. In the approach described by the guide, skilled professionals—averaging about five years of relevant experience—complete tasks across areas including software engineering, machine learning, and cybersecurity. METR records their working time, then gives AI agents multiple attempts at the same tasks.
The guide says METR runs six independent attempts per task and checks transcripts for cheating, counting it as failure. Researchers then fit a curve showing how success changes with task length. The 80% horizon is the point where the curve reaches four successful attempts out of five; the 50% horizon is where success is estimated at one in two. The metric reflects task difficulty rather than the time the AI itself spends working, and agents can complete successful tasks faster than humans.
METR’s results show a sharp rise across the listed releases: from under a second for GPT-2 in 2019 to 66 minutes for GPT-5.2 in December 2025, then to the Mythos Preview result in April 2026. The guide also cites METR’s February–March 2026 Frontier Risk Report, which estimated the labs’ internal frontier at 80% reliability as at least three hours but under four hours. That estimate is separate from the public benchmark record.
“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 185.9-minute estimate has a broad confidence interval, from 97.3 to 398.5 minutes, so the precise horizon is uncertain. METR also says the public results page is no longer actively updated; the record cited in the guide is from April 2026, and the page’s last update was September 8, 2026. The figures do not establish how models released or tested after that period perform.
It is also unclear how well the benchmark maps to any particular team’s work. The guide cites a METR comparison in which messier tasks cut the measured horizon by almost a factor of three, but the supplied material does not specify enough detail to generalize that reduction across roles or industries. A benchmark success rate is an estimate across tested tasks, not a guarantee for an individual ticket, and an 80% horizon is not 100% reliability.
Teams Need Their Own Tests
The guide recommends that teams begin with work measured in hours and defined by clear, checkable outcomes, then track completion and review effort on their own tasks. It also advises choosing a model for each type of work and repeating those measurements quarterly. These steps would give teams evidence about whether published benchmark results match their codebases and processes.
For the broader field, the next useful development would be updated, public measurements that allow comparisons with the April 2026 record. The source material does not give a date for a new METR results release or say when the cited public benchmark will resume updates. Until newer data are available, teams should treat the roughly three-hour figure as a qualified benchmark result, not a current guarantee of independent work.
Key Questions
Does a three-hour AI horizon mean the agent can work alone for three hours?
No. The horizon measures task difficulty using the time a skilled human would need. It does not measure how long the AI runs independently or mean that supervision is unnecessary.
What does an 80% time horizon mean?
It is the estimated task length at which the agent succeeds four times out of five, based on METR’s fitted success curve. It is not a guarantee that a particular task will succeed.
Which model has the highest result in the cited public data?
The guide identifies an early Claude Mythos Preview, dated April 7, 2026 in METR’s results file, at 185.9 minutes for the 80% horizon. Its confidence interval is wide, and the result does not cover later releases.
Why is the 50% horizon much longer?
A 50% horizon accepts a lower success rate than an 80% horizon, allowing the estimate to extend to longer tasks. For the model cited, the 50% figure is about 17.4 hours, but METR warns that measurements above 16 hours are unreliable with its current task suite.
How should a team apply the benchmark?
Use it as a reference, then test AI on the team’s own small, clearly specified tasks. Record success, review time, and errors, and keep a human check where failures would matter.
Source: ThorstenMeyerAI.com
