🔍 Read the full analysis: How Long Can AI Stay On Task? The Time Horizon Explained 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 benchmark records an 80% time horizon of 185.9 minutes for an early Claude Mythos Preview, based on results dated April 7, 2026. The measure estimates the length of tasks AI agents can complete at a given success rate; it is not a promise that an agent will work autonomously for that long. METR’s public results are no longer actively updated, and performance can fall on messier tasks.
METR’s public benchmark puts the 80% task horizon for an early version of Anthropic’s Claude Mythos Preview at 185.9 minutes, or about three hours, according to a results file dated April 7, 2026. The figure estimates the length of tasks the model can complete reliably under METR’s test conditions; it does not mean the AI can work unattended for three hours or succeed on every task of that length.
METR, an independent research nonprofit, measures tasks by how long skilled human experts take to complete them, not by how much time the AI spends. Its current Time Horizon 1.1 suite contains 228 tasks, spanning software engineering, machine learning and cybersecurity. METR says the professionals who worked on the tasks had about five years of relevant experience on average.
For each task, METR runs an AI agent six times independently and checks transcripts for cheating, which counts as failure. It then estimates how success rates change with task length. The time horizon is the point where the fitted success curve reaches a selected reliability level. At the 80% threshold, the Mythos Preview result is 185.9 minutes; its reported confidence interval is wide, from 97.3 to 398.5 minutes.
The benchmark’s record 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 reaching 185.9 minutes for Mythos Preview. Those results describe particular model versions and METR tasks, not all AI systems or workplace assignments. METR’s public page says it is no longer actively updated and was last updated September 8, 2026.
What the Three-Hour Estimate Means for Planning
The horizon measures performance in relation to task duration and success rate. Teams can use it as one reference when assessing which tasks might be suitable for delegation and planning human review. The roughly three-hour result indicates that, under comparable conditions, the leading agent in this benchmark reached an estimated 80% success rate on tasks taking human experts about three hours.
An 80% threshold corresponds to an estimated failure rate of one in five attempts across the measured tasks. The benchmark does not establish that a model will perform equally well on a company’s own code, data or procedures. The source article cites a METR trial in which experienced developers were 19% slower with AI while believing they were 20% faster, and a separate study where reviewers caught 31% of errors that required task-specific rules to identify. These findings concern AI-assisted work and error review, rather than the benchmark’s task horizon.
The source recommends dividing work into hour-scale pieces with clear acceptance checks, starting with testable tasks, and measuring results against a team’s own work. Local evaluations can help determine whether performance on the benchmark transfers to a particular workflow.
Popular right now

Ninja Foodi Smart XL 6-in-1 Indoor Grill and Air Fryer with Thermometer, 4 qt, Black and Silver (FG551)

Roborock Qrevo S Pro Robot Vacuum and Mop (2026)

Vitamix 5-Speed Immersion Blender, 625-Watt, Stainless Steel/Black
As an affiliate, we earn on qualifying purchases.
How METR Measures Agent Task Length
METR’s measure is based on the human working time needed for a task. The organization plots the agent’s success across tasks of different lengths and estimates where reliability falls to a chosen threshold. A 50% horizon and an 80% horizon therefore answer different questions: the first marks a coin-flip level, while the second requires success on four out of five attempts.
The same Mythos Preview result illustrates the difference. METR’s file gives it a 50% horizon of 1,044.8 minutes, about 17.4 hours, compared with 185.9 minutes at 80%. METR warns that measurements above 16 hours are unreliable with the current task suite, making the longer estimate especially uncertain. METR also says completed tasks are typically performed by agents several times faster than humans; the horizon concerns task difficulty, not the duration of the agent’s work.
Task conditions matter too. The source reports that a METR comparison found messier tasks cut the horizon by almost a factor of three. In a February–March 2026 Frontier Risk Report, METR estimated the labs’ internal frontier at 80% to be at least three hours but under four hours. These are benchmark-specific findings, not evidence that every agent can reliably handle tasks of comparable length in ordinary workplaces.
“a measure of the difficulty of a task, rather than the time an AI spends to complete the task.”
— METR, describing the interpretation of its measure
Limits of the Public Benchmark Record
The 185.9-minute estimate has a confidence interval spanning roughly 1.6 to 6.6 hours, indicating substantial statistical uncertainty. METR’s public results page has not been actively updated since September 8, 2026, and the supplied results end with a model entry dated April 7, 2026. It is therefore unclear from this public record whether newer systems have longer horizons.
The task suite is also cleaner and more clearly specified than many real jobs, according to the source material. Results may not transfer to work involving ambiguous requests, changing requirements, unfamiliar internal tools or costly mistakes. METR’s warning about measurements above 16 hours also makes the 50% estimate of 17.4 hours less dependable. The figures do not establish success rates for any particular organization’s workload.
Measure Results on Your Own Work
The source recommends that teams test agents on bounded, verifiable tasks, record both completion and checking time, and repeat those evaluations on their own work each quarter. The benchmark can provide a reference point for task length, while local results can help assess an agent’s performance on a specific workflow.
No next public METR results update is specified in the source material. Until a newer benchmark or an updated task suite is published, the April 2026 Mythos Preview result remains the latest record cited here, rather than a current measure of every leading model.
Key Questions
What does an AI time horizon measure?
It estimates the length of tasks, measured by human expert completion time, that an AI agent can complete at a chosen reliability level. It does not measure how long the agent runs.
What is the best measured 80% horizon in the cited results?
METR’s results file lists 185.9 minutes for an early Claude Mythos Preview dated April 7, 2026. The reported confidence interval is 97.3 to 398.5 minutes.
Does a three-hour horizon mean an agent will work alone for three hours?
No. It refers to the estimated human duration of tasks where the model reaches a particular success rate. The AI’s own completion time can differ, and the result is not a guarantee for an individual task.
Why is the 50% horizon much longer than the 80% horizon?
A lower reliability threshold includes longer tasks. For the cited Mythos Preview result, the 50% horizon is about 17.4 hours, compared with about 3.1 hours at 80%; METR cautions that measurements above 16 hours are unreliable with its current task suite.
Can companies apply the benchmark directly to their own work?
Not as a guarantee. METR’s tasks are controlled and well specified, while workplace assignments can be messier. Teams should test their own tasks, define acceptance checks and account for human review time.
Source: ThorstenMeyerAI.com
