🔍 Read the full analysis: A Practical Framework For Measuring AI’s Time Horizon 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
A November 18, 2026 guide from ThorstenMeyerAI.com explains how METR measures AI time horizons: the length of tasks, timed by skilled human work, that an AI agent can complete at a specified reliability. The guide reports a public 80% benchmark of 185.9 minutes for an early Claude Mythos Preview, but cautions that real-world task messiness, reliability requirements and human review can narrow the practical value.
ThorstenMeyerAI.com published a guide on November 18 explaining how teams can use METR’s AI time horizon to judge which tasks to assign to AI agents. The guide cites an 80% time horizon of 185.9 minutes for an early Claude Mythos Preview, while stressing that the benchmark is not a guarantee of success on a company’s own work.
METR, an independent research nonprofit, defines a time horizon by the human expert time needed for a task and the reliability with which an AI agent completes tasks of that length. The metric does not mean the AI spends that long working. According to the guide, agents typically complete successful benchmark tasks faster than human experts.
The guide says METR’s Time Horizon 1.1 suite contains 228 tasks spanning software engineering, machine learning and cybersecurity. Skilled professionals first complete the tasks so researchers can record human working time. METR then runs agents six times per task, checks transcripts for cheating, and estimates the task length at which success falls to a chosen threshold, such as 80% or 50%.
METR’s cited results put the early Claude Mythos Preview, dated April 7, 2026, at 185.9 minutes at 80% reliability, or about three hours and six minutes. The guide reports a confidence interval of 97.3 to 398.5 minutes. It also lists earlier points in the series: GPT-4 at 0.89 minutes in March 2023, o3 at 29.98 minutes in April 2025, and GPT-5.2 at 66 minutes in December 2025. These are benchmark results, not measurements of every task or deployment.
Turning Benchmarks Into Work Plans
For managers, the measure offers a more useful planning 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? That framing can help teams divide projects into units with clear acceptance tests, assign more testable work first, and set aside time for review.
The guide warns against treating a horizon as an operating guarantee. The reported result describes a success rate across benchmark attempts; it does not mean any individual three-hour task has an 80% chance of success in every workplace. Teams still need to account for their code, documentation, security requirements and cost of failure. The practical implication is to measure performance on local tasks and budget for checking rather than assume benchmark capability transfers unchanged.
Reliability requirements change the apparent horizon. The guide reports that the same Mythos Preview model has a 50% horizon of 1,044.8 minutes, compared with 185.9 minutes at 80%. METR cautions that measurements above 16 hours are unreliable with its current task suite, so the longer figure should be treated carefully. A lower success threshold allows longer tasks into the measure, but may be too weak for work that needs dependable completion.
Products Worth Considering

Dell UltraSharp U2424H 24-Inch FHD Monitor

Samsung ViewFinity S7 27-Inch 4K UHD Monitor

FLEXISPOT Oversized Outdoor Rocking Chair with Thick Cushion, Grey
As an affiliate, we earn on qualifying purchases.
How METR Builds the Measure
The guide describes the time horizon as a fitted estimate rather than a direct claim that an agent can reliably finish every task of a certain duration. Researchers compare task success with human-measured duration, then locate where the estimated success curve crosses a selected reliability level. The 80% version means roughly four successful attempts in five at that point; the 50% version marks a less reliable threshold.
The cited public results show a rise from under a second for GPT-2 in 2019 to the 185.9-minute Mythos Preview result in April 2026. The guide says its own trend fits suggest the line has doubled every three to seven months, depending on the data selected. It projects an eight-hour workday at 80% sometime between September 2026 and November 2027, but that is a forecast based on past trends, not a confirmed capability. The source says nothing after April 2026 had been publicly measured in the cited record.
METR’s task design also limits how broadly the number can be applied. According to the guide, benchmark tasks are comparatively clean and well specified. In one METR comparison, messier tasks reduced the horizon by almost a factor of three. That difference matters for workplace backlogs, where requirements may be incomplete, dependencies hidden or acceptance criteria disputed.
“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 cited record has important limits. Its confidence interval spans about 1.6 to 6.6 hours for the Mythos Preview’s 80% result, and the guide says METR’s public results page is no longer actively updated, with its last update on September 8, 2026. The source also reports that METR estimated the labs’ internal frontier at “≥3h but <4h” at 80% in a February–March 2026 risk report; that estimate is separate from the public benchmark result.
It is not clear from the source when METR will resume public updates, how the cited model performs on any particular company’s workload, or whether the projected eight-hour horizon will be reached. The guide also describes evidence that oversight can fail: it reports a 2025 trial in which experienced developers were 19% slower with AI despite believing they were 20% faster, and a separate study where reviewers caught 31% of errors that required task rules to identify. Those figures are reported by the guide and do not establish that every team will see the same results.
Test Against Your Own Backlog
The guide recommends that teams start with small, clearly specified tasks and objective acceptance checks, then expand only when results support it. It advises assigning AI the testable portions of work, choosing models for the task rather than by headline benchmark, and explicitly budgeting for human review and correction.
Its proposed next step is to repeat evaluations on a company’s own work every quarter. That would let teams compare success rates across task lengths and detect whether changes in models or workflows improve practical performance. A further public METR update, if one is issued, would also help clarify how the benchmark has moved since the April 2026 result; the source does not give a date for any such update.
Key Questions
What does an AI time horizon measure?
It estimates the human-expert duration of tasks an AI agent can complete at a specified reliability level. It does not measure how long the agent itself takes to work.
What is the latest public result cited in the guide?
The guide cites an 80% horizon of 185.9 minutes for an early Claude Mythos Preview, dated April 7, 2026 in METR’s results file. Its reported confidence interval is 97.3 to 398.5 minutes.
Does a three-hour horizon mean every three-hour task will succeed?
No. It is a benchmark estimate at a stated success threshold, not a promise about an individual task. Performance may differ on a team’s own work, especially when tasks are less clearly specified.
Why does the success-rate threshold matter?
A lower threshold includes longer tasks but allows more failures. The guide reports a 50% horizon of about 17.4 hours for the Mythos Preview, compared with about 3.1 hours at 80%, and notes METR’s warning that measurements above 16 hours are unreliable with its current suite.
How should teams apply the measure?
Start with short tasks that have clear acceptance tests, measure outcomes on the team’s own workload, and reserve time for review. The guide recommends repeating those evaluations quarterly rather than relying on a single public benchmark.
Source: ThorstenMeyerAI.com
