🔍 Read the full analysis: Use Time Horizons To Make Sense Of What AI Can Do Today 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 results list an 80% task-completion horizon of 185.9 minutes for an early Claude Mythos Preview released in April 2026. That is a measure of task difficulty in skilled human work time, not how long the AI takes to work, and it does not guarantee success on any individual task. The result is based on relatively well-specified tasks; reliability, task messiness and review effort all affect how useful it is for planning.
A November 18 guide from ThorstenMeyerAI.com argues that teams should judge AI systems by the length of tasks they can complete reliably, rather than by whether they can perform a task at all. Citing METR, an independent research nonprofit, the guide says the best result in the cited public data is an 80% time horizon of 185.9 minutes for an early Claude Mythos Preview released in April 2026. The measure offers a practical way to discuss AI task capability, but it is not a guarantee for a specific job or workplace.
METR defines a time horizon as the duration of a task, measured by the time skilled human experts take, at which an AI agent is predicted to succeed at a specified reliability level. The guide uses the 80% threshold: across repeated attempts on tasks of that length, the model is predicted to succeed four times out of five. That human-time figure describes task difficulty; it does not mean the AI spends three hours completing the work. METR says agents that finish tasks typically do so faster than people.
The source says METR’s Time Horizon 1.1 suite contains 228 tasks spanning seconds to more than eight hours. Skilled professionals, with an average of about five years of relevant experience, completed tasks that were timed. METR then ran each task six times with an AI system, treated detected cheating as failure, and fitted a curve showing how success changes with task length. The 80% horizon is the point where that curve crosses the four-in-five success rate.
In the results file cited by the guide, the 80% horizon 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. The cited record is 185.9 minutes for the early Mythos Preview, dated April 7, 2026. METR’s estimate has a wide confidence interval, from 97.3 to 398.5 minutes. The source also notes that METR’s public results page was no longer being actively updated and was last updated September 8, 2026.
Using the Horizon to Plan Work
The measure gives managers a more useful starting point than a broad question such as whether AI can code, analyze data or complete a project. It frames capability around task length and reliability, helping teams identify work that can be divided into bounded pieces with clear checks. A short, testable ticket may be suitable for an AI-assisted workflow even when an entire project is not.
But an 80% result leaves a meaningful failure rate: one in five attempts may not succeed at the benchmark’s crossing point. Teams need to budget time for checking, correction and escalation rather than counting every attempted task as completed work. The guide recommends starting with testable tasks, setting acceptance criteria, choosing models for specific jobs and measuring performance on a team’s own work at regular intervals.
The source also cites a 2025 METR trial in which experienced developers using AI were 19% slower while believing they were 20% faster. It reports that reviewers in another study caught only 31% of errors that required applying the task’s rules. Those figures do not establish that every team will see the same result, but they underline why model capability and net productivity are different questions.
Products Worth Considering

Jackery Explorer 2000 v2 Portable Power Station, 2042Wh, 2200W

Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W

Bose QuietComfort Wireless Noise Cancelling Headphones, Sandstone
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Length
The time-horizon approach is intended to resolve a mismatch between headline demonstrations and work in existing organizations. A model may generate a working application from a brief prompt yet struggle with a small change in an unfamiliar codebase. METR’s framing asks how long a skilled person would need for the task and how reliably an agent completes tasks of comparable length.
Reliability changes the apparent horizon. For the same Mythos Preview, the guide gives a 50% horizon of 1,044.8 minutes, or about 17.4 hours, compared with 185.9 minutes at 80%. The longer figure is a much weaker planning threshold: it represents a one-in-two success rate. METR also cautions that measurements above 16 hours are unreliable with its current task suite, making the 17.4-hour estimate especially uncertain.
Task conditions matter as well. The source says METR’s benchmark tasks are clean and well specified, and that one METR comparison found messier tasks reduced the horizon by almost a factor of three. The guide’s own trend fits suggest the horizon has doubled every three to seven months, but that is an extrapolation from past results, not a guarantee. Its estimates for when a model might reach a full eight-hour workday at 80% range from September 2026 to November 2027, depending on which years are used.
“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 Estimate
The cited 185.9-minute estimate is not a guarantee that a model will finish any particular three-hour task, and its confidence interval is broad. The result describes performance across METR’s benchmark and depends on the task mix, evaluation method and chosen reliability threshold. The source says there is no public measurement after April 2026 in the results it reviewed; it does not establish what current internal or later model evaluations might show.
It is also unclear how well the benchmark result transfers to each company’s work. Real tickets can involve incomplete requirements, unfamiliar systems, coordination or judgment calls that are not captured by cleanly specified benchmark tasks. The guide’s forecast of continued rapid growth is a projection based on historical fits, not a confirmed timetable. Its discussion does not provide a new independent evaluation of model performance.
Testing AI on Team Work
The guide recommends that teams apply the measure cautiously: break projects into small tasks with clear acceptance checks, start with work that can be tested, and include human review in estimates of time and cost. It also suggests selecting a model for each task type and tracking outcomes on the organization’s own work at least quarterly.
For a clearer view of how capabilities are changing, the next useful development would be updated, public benchmark results and comparable evaluations on less structured tasks. Until those are available, the cited three-hour horizon is a reference point for METR’s task suite—not a promise of what an AI system can deliver across a full working day.
Key Questions
What does an AI time horizon measure?
It estimates the length of a task, measured by skilled human completion time, at which an AI agent reaches a specified success rate. At the 80% threshold, the model is predicted to succeed on four out of five attempts.
Does a three-hour horizon mean the AI works for three hours?
No. The three hours refer to how long skilled people take to complete tasks of that difficulty. The figure does not describe the AI’s working time.
What is the latest public result cited in the guide?
The guide cites METR’s April 7, 2026 result for an early Claude Mythos Preview: an 80% horizon of 185.9 minutes, with a confidence interval from 97.3 to 398.5 minutes.
Why is the 50% horizon much longer?
A 50% threshold requires success on only half of attempts, while an 80% threshold requires four successes in five. The guide cites about 17.4 hours at 50% for the same model, but says METR considers measurements above 16 hours unreliable with its current task suite.
Can companies use the result to predict productivity?
Not by itself. The benchmark describes performance on METR tasks, and workplace results may differ with task clarity, review needs and the systems involved. Teams should test AI on their own work and account for checking and correction.
Source: ThorstenMeyerAI.com
