🔍 Read the full analysis: How Teams Can Gauge What AI Can Do Today 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 lists an early Claude Mythos Preview, measured in April 2026, with an 80% task-duration horizon of 185.9 minutes. That is a benchmark result, not a promise that AI can reliably complete any three-hour job; task messiness, checking costs and the benchmark’s limits matter.
Teams deciding which work to assign to AI can use a benchmark from research nonprofit METR that measures how long a skilled person would take to complete tasks AI agents can perform reliably. A guide published by ThorstenMeyerAI.com on November 18 says the best model in METR’s cited public results, an early Claude Mythos Preview, had an 80% task-duration horizon of 185.9 minutes—about three hours—on the benchmark’s tasks.
METR’s measure is based on human working time, not how long the AI takes to run. It estimates the task length at which an AI agent is predicted to succeed at a specified rate. The guide focuses on the 80% threshold: four successful attempts out of five. METR’s benchmark suite, Time Horizon 1.1, contains 228 tasks. Skilled professionals complete the tasks to establish human-time estimates; AI agents then make multiple attempts, with METR reviewing transcripts for cheating.
The results file cited by the guide puts the early Claude Mythos Preview, dated April 7, 2026, at 185.9 minutes at the 80% threshold. Its reported confidence interval is broad, from 97.3 to 398.5 minutes. For comparison, the same results list 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 a ranking of performance on every kind of workplace task.
The success threshold changes the estimate. For the same Mythos Preview model, the guide reports a 50% horizon of 1,044.8 minutes, or about 17.4 hours, while noting that METR warns measurements above 16 hours are unreliable with its current task suite. In addition, the guide says METR’s public results page was no longer actively updated as of its last update on September 8, 2026. The cited record therefore does not establish how models released or measured after April performed.
Turning Benchmark Results Into Work Plans
The measure gives team leads a more practical question than whether AI “can do” a task: how long would a skilled person need, and at what reliability does the agent complete comparable work? That can help teams decide which tasks to pilot, how much human review to budget and where to divide larger assignments into smaller pieces with clear acceptance checks.
But the number is not a guarantee for a specific ticket. METR’s tasks are described as relatively clean and well specified, while the guide cites a METR comparison in which messier tasks reduced the horizon by almost a factor of three. A team applying the benchmark to its own code, data or internal processes should treat the headline horizon as a starting point, not a delivery estimate. Review time, failure recovery and integration work also affect whether using AI saves time overall.
The guide also points to research that cautions against assuming that assistance automatically makes people faster. It reports that in a 2025 METR trial, experienced developers working with AI were 19% slower, even though they believed they were 20% faster. That finding concerns the trial described in the source, not every team or workflow, but it underlines why local measurement matters.
Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

DJI Mini 4K Fly More Combo Drone with 4K Camera and 3 Batteries

Logitech MX Mechanical Mini for Mac Wireless Keyboard, Space Gray
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Duration
METR’s approach starts by timing skilled people doing benchmark tasks in fields including software engineering, machine learning and cybersecurity. The guide says the contractors had an average of about five years of relevant experience. METR then has AI agents attempt tasks multiple times—in the described method, six independent runs per task—and evaluates success. Plotting success against human task duration produces an estimate of the task length associated with a chosen reliability threshold.
A three-hour horizon does not mean the model works for three hours, nor that every task taking a person three hours will be completed. It describes estimated task difficulty at a particular success rate. The guide says agents that succeed typically complete tasks several times faster than people, but speed on successful attempts does not remove the risk of failure or the need to check the result.
The source presents a possible longer-term trend, saying its own fits suggest the measured horizon has doubled every three to seven months, if that pattern continues. It estimates an eight-hour workday at the 80% threshold could fall between September 2026 and November 2027, depending on which years are used. Those dates are projections in the guide, not measured outcomes; its cited public data stop at the April 2026 result.
“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 result has a wide reported confidence interval, and METR’s cited page was no longer being actively updated, according to the guide. It is unclear from the source whether newer models have since been evaluated using a comparable public method. The guide also does not establish how closely its benchmark tasks match any particular employer’s work; differences in specifications, tools, data access and review standards could change results.
Other figures need similar care. The 50% estimate exceeds the range METR says is reliable with its current task suite, while projections about reaching an eight-hour horizon depend on trend fits and may not hold. The source also references a separate reviewer study in which reviewers caught 31% of errors requiring task-specific rules to identify, but does not provide enough study detail here to generalize that figure across teams. Performance on a benchmark cannot establish reliability in a team’s own workflow without testing that workflow.
Test the Work Your Team Has
The guide recommends breaking assignments into hour-scale pieces with explicit acceptance checks, starting with work that can be tested objectively, and budgeting time for human review. It also advises selecting models by task rather than assuming one model is best for every assignment, then measuring outcomes on the team’s own work every quarter.
For teams, the next useful step is a controlled pilot: record how long comparable tasks take people, define what counts as a correct result, and track AI success, review effort and rework separately. The source does not identify a future date for METR to resume public updates. Until comparable new measurements are available, the April 2026 result and the guide’s projections should be treated as dated evidence rather than a current guarantee.
Key Questions
What does an AI time horizon measure?
It estimates the length of a task, measured by the time skilled people take, at which an AI agent is predicted to succeed at a chosen reliability level. It does not measure how long the AI itself works.
What does an 80% horizon mean?
It means the agent is estimated to complete tasks of that duration successfully in four out of five attempts under the benchmark conditions. It is not a promise about any particular task.
What was the best public result cited in the guide?
The guide reports that an early Claude Mythos Preview had an 80% horizon of 185.9 minutes in METR’s results file, dated April 7, 2026. Its confidence interval was 97.3 to 398.5 minutes, and the result reflects METR benchmark tasks.
Can teams apply that three-hour figure directly to their projects?
No. The benchmark is a reference point, and the guide says messier tasks reduced the horizon by almost a factor of three in a METR comparison. Teams should test their own work, define acceptance criteria and include review and rework in their estimates.
Source: ThorstenMeyerAI.com
