🔍 Read the full analysis: Could AI Handle Your Next Long Task? Start With The Time Horizon on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s public benchmark puts an early Claude Mythos Preview at a 185.9-minute time horizon with an 80% success rate, based on tasks timed for skilled human workers. The measure is not a guarantee: reliability drops on longer tasks, benchmark work is cleaner than many workplace assignments, and the public record has not been updated since September 2026.
METR’s public benchmark puts an early Claude Mythos Preview at an 80% time horizon of 185.9 minutes, or about three hours, on tasks measured by the time skilled humans need to complete them. The result, dated April 7, 2026, offers teams a way to think about which work AI agents may handle, but it is not a promise that a model will reliably finish any three-hour workplace assignment.
The time horizon is a measure of task difficulty in human working time, not how long the AI itself takes. METR defines it as the task duration at which an agent is predicted to succeed at a specified reliability level. In the 80% version, the agent must succeed on roughly four out of five attempts at the measured task length.
METR’s Time Horizon 1.1 suite contains 228 tasks in software engineering, machine learning and cybersecurity. The nonprofit says it uses skilled professionals, averaging about five years of relevant experience, to establish human completion times. AI agents make six independent attempts per task; METR reviews run transcripts and counts cheating as failure, according to the source article.
The public results show the 80% 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. The April 2026 Mythos Preview result has a wide reported confidence interval, from 97.3 to 398.5 minutes. METR’s public results page was marked no longer actively updated and was last updated September 8, 2026, the article says.
How Teams Can Use the Three-Hour Measure
The benchmark can help managers frame work by task length and verifiability rather than asking broadly whether AI can do a job. A ticket that a skilled worker completes in a few hours, with a clear test or acceptance check, is a more useful candidate for a controlled trial than an open-ended assignment with ambiguous requirements.
But the 80% figure also means roughly one failure in five at the measured level, not dependable completion without oversight. Teams planning around the result need to include time for review, error correction and escalation. The benchmark does not establish the success rate on a company’s own codebase, tools or processes; those need to be tested separately.
The source article reports that a separate METR comparison found messier tasks reduced the measured horizon by nearly a factor of three. It also cites a 2025 trial in which experienced developers took 19% longer with AI while estimating they were 20% faster. Those results underline that perceived speed and benchmark performance do not by themselves show that AI improves a team’s total output.
Products Worth Considering

Canon SELPHY QX20 Compact Photo Printer – Dark Gray

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

Logitech MX Creative Console with 9 Customizable Keys, Graphite
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Length
METR estimates a horizon by connecting human completion times with repeated AI-agent attempts. Shorter tasks tend to have higher success rates, while success falls as task length rises. METR fits a curve to those results and identifies where it reaches a chosen reliability level, such as 80% or 50%.
The source article emphasizes that the number is not an estimate of the AI’s working speed. METR says agents that succeed on tasks are typically several times faster than humans, but that does not change the definition: the horizon describes the difficulty of work using a human-time scale. The article uses the 80% measure for planning and contrasts it with the Mythos Preview’s 50% horizon of 1,044.8 minutes, about 17.4 hours. METR cautions that measurements above 16 hours are unreliable with its current task suite.
For comparison, METR’s February–March 2026 Frontier Risk Report estimated the labs’ internal frontier at at least three hours but under four hours at 80%, according to the article. The source also gives its own forecast for an eight-hour 80% horizon, ranging from September 2026 to November 2027 depending on which years of trend data are used. That is a forecast, not a measured result.
Where the Benchmark Falls Short
The public record does not show how the leading model would perform on ordinary company work, which may involve incomplete instructions, legacy systems, coordination or changing requirements. The source article says messier tasks can shorten the measured horizon, but the material provided does not specify enough about that comparison to apply a precise adjustment to a particular team’s work.
The Mythos Preview estimate also has a broad confidence interval, and the 50% result extends beyond the task suite’s stated reliable measurement range. The source says no public measurements after April 2026 were available at the time of publication. It does not establish whether newer models have longer horizons, or whether the same gains would carry over to less structured tasks.
Finally, an 80% horizon is a fitted benchmark measure, not a guarantee for every task at that length. The source article cites a separate review study in which reviewers caught only 31% of errors that required knowledge of the task’s rules to identify. The details supplied do not explain the study design or its sample, so that figure should not be generalized beyond the cited finding.
Testing AI on Real Team Work
The practical next step is for teams to test agents on their own representative tasks, record completion and error rates, and repeat the evaluation as models and workflows change. The source article recommends breaking work into pieces measured in hours, starting with tasks that have clear acceptance checks, and budgeting for human review rather than treating the benchmark horizon as a handoff threshold.
It also recommends choosing models by task and re-measuring performance every quarter. METR’s public page was not actively updated as of the date reported in the source, so a newer public benchmark result would help establish whether the April 2026 record has changed. Until then, an eight-hour horizon remains a forecast range in the article, not an observed capability.
Key Questions
What does a three-hour AI time horizon mean?
It means the benchmark estimates that the model reaches a specified success rate on tasks that skilled humans take about three hours to complete. It does not mean the AI works for three hours or will finish every such task.
Which model holds the reported 80% result?
The source article identifies an early Claude Mythos Preview, dated April 7, 2026 in METR’s results file, with an estimated 185.9-minute horizon at 80% success. The reported confidence interval is 97.3 to 398.5 minutes.
Why is the 80% result different from the 50% horizon?
A model can reach a 50% success rate on longer tasks than it can complete reliably four times out of five. The article gives Mythos Preview’s 50% horizon as about 17.4 hours, while warning that METR considers measurements above 16 hours unreliable with its current task suite.
Can a team use the benchmark to decide what to automate?
It can inform a trial, but it cannot establish performance on a team’s own work. The source article recommends starting with bounded tasks that have clear acceptance checks, measuring results locally and allowing time for human review.
Source: ThorstenMeyerAI.com
