AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: AI For Teams: Understanding What It Can Do And For How Long on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

  • Fast, free delivery for dorm and study essentials
  • Prime Video and Amazon Music included
  • Member-only deals
Try Prime for Young Adults Free trial for eligible 18–24 year olds
As an affiliate, we earn on qualifying purchases.

TL;DR

METR’s public benchmark lists an 80% time horizon of 185.9 minutes for an early Claude Mythos Preview released in April 2026. The measure estimates the length of tasks, by skilled human working time, that an AI agent completes with a specified reliability; it is not a measure of how long the AI takes to work. METR’s public results page is no longer actively updated, and task complexity and review needs limit how directly the figure applies to workplace projects.

A November 18 guide from ThorstenMeyerAI.com explains how teams can use METR’s task-time benchmark to estimate which work AI agents may complete reliably. METR’s public results file lists an 80% time horizon of 185.9 minutes for an early Claude Mythos Preview released in April 2026: an estimate that the model succeeds on tasks taking skilled people about three hours, not that it works for three hours or succeeds on every such task.

METR, an independent research nonprofit, defines a time horizon as the duration of a task, measured by expert human completion time, at which an AI agent is predicted to succeed with a specified reliability. Its current Time Horizon 1.1 benchmark includes 228 tasks in areas including software engineering, machine learning and cybersecurity. METR says the human professionals completing the tasks had an average of about five years of relevant experience.

For each task, METR runs an AI agent six times independently and checks the transcripts for cheating, which counts as failure. It then estimates how success rates change with task length. The reported horizon is the point where the fitted curve crosses a chosen success rate: the guide focuses on 80% reliability, or four successful attempts out of five.

METR’s results file shows the 80% figure rising from 0.89 minutes for GPT-4 in March 2023 to 29.98 minutes for o3 in April 2025, then 66 minutes for GPT-5.2 in December 2025. Its listed record is 185.9 minutes for the early Mythos Preview, dated April 7, 2026. The estimate has a wide confidence interval, from 97.3 to 398.5 minutes. METR’s public page says it is no longer actively updated; its stated last update was September 8, 2026.

At a glance
reportWhen: Guide published November 18, 2026; cite…
The developmentA guide published November 18, 2026, uses METR’s task-time benchmark to explain what AI agents can reliably complete and how teams should interpret the measure.

Planning Work Around Reliability

The measure gives team leads a more useful planning question than whether AI “can” do a task: how long would an expert need, and how often does the system complete comparable work? That can help teams decide which work to trial, how to divide projects, and how much human review to budget. The guide recommends starting with short, testable tasks, using clear acceptance checks and selecting tools by task rather than assuming one model fits every job.

The 80% figure is not a guarantee for an individual task. Four successes in five leaves a failure on the fifth attempt on average, and a team still needs someone able to detect and correct errors. The source guide cites a 2025 METR trial in which experienced developers took 19% longer with AI while believing they were 20% faster. It also cites a separate study where reviewers caught 31% of errors that required the task’s rules to identify. These findings caution against treating apparent speed or confidence as proof of net productivity.

METR’s benchmark can inform decisions, but it does not replace measurement on a team’s own work. Internal code, ambiguous requests, dependencies and consequences of mistakes may differ from benchmark tasks. Teams need to track completion quality and review effort, not just whether a model produces an answer.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

How METR Measures Task Length

The time horizon is based on human working time, not the duration of an AI session. METR notes that agents are typically several times faster than humans on tasks they complete, but speed is separate from task difficulty and reliability. A three-hour horizon means the relevant task is estimated to require a skilled person about three hours; it does not say the agent needs that much time.

The selected success threshold also changes the result. For the early Mythos Preview, METR’s file gives a 50% horizon of 1,044.8 minutes, about 17.4 hours, compared with 185.9 minutes at 80%. The guide says that longer estimate is less reliable because METR warns that measurements above 16 hours are unreliable with its current task suite. The difference shows why a horizon should always be reported with its reliability threshold.

Benchmark conditions matter too. The guide reports that one METR comparison found messier tasks reduced the horizon by almost a factor of three. It also relays the author’s own trend fits suggesting the time horizon has doubled every three to seven months, depending on the period analyzed. That is a retrospective estimate, not a guarantee of future progress. The guide’s projections for an eight-hour workday range from September 2026 to November 2027, but it says no public measurements after April 2026 were available for its account.

“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

Benchmark Limits and Open Questions

The reported record is an estimate, not a promise that a model will complete any particular three-hour task. Its confidence interval spans roughly 1.6 to 6.6 hours, and the benchmark’s task suite cannot capture every workplace condition. The source guide reports that messy tasks can shorten the measured horizon, while the 50% estimate above 16 hours carries an explicit reliability warning from METR.

It is also unclear how the benchmark would compare with current systems after April 2026: the guide says no later public measurements were available, and METR’s public results page is no longer actively updated. The source does not establish whether unpublished evaluations exist or how the measured results translate to a particular company’s code, processes or risk tolerance. Forecasts based on past growth rates remain projections, not measured outcomes.

Testing AI on Team Backlogs

The guide recommends that teams begin with work that can be divided into tasks lasting hours and checked against clear acceptance criteria. They should assign AI the testable parts first, budget time for review, choose models according to the task and compare results against their own baseline. Repeating those measurements quarterly would help teams see whether performance is changing on their actual work.

For managers, the next useful evidence is not simply a higher benchmark record. It is whether an AI system completes a team’s tasks at the required quality, how much checking people must do, and whether the full process saves time. METR’s public benchmark provides one reference point; the source guide does not report a new METR release or a scheduled update.

Key Questions

What does an AI time horizon measure?

It estimates the length of a task, measured by skilled human working time, that an AI agent completes at a specified success rate. It does not measure how long the AI itself takes.

What is the latest public 80% figure cited in the guide?

METR’s results file lists 185.9 minutes for an early Claude Mythos Preview dated April 7, 2026. The estimate has a broad confidence interval, and METR’s public results page is no longer actively updated.

Does an 80% horizon mean a team can skip review?

No. It means the model is estimated to succeed on four out of five attempts at that task length under the benchmark conditions. Teams still need acceptance checks and human review, particularly where errors carry consequences.

Why is the 50% horizon much longer?

A lower success threshold allows longer tasks into the estimate. For the early Mythos Preview, the listed 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.

How should teams apply the benchmark?

Use it as a reference, then test AI on the team’s own work. The guide recommends breaking work into clear, testable tasks, budgeting review time, choosing tools by task and measuring results on a regular schedule.

Source: ThorstenMeyerAI.com

You May Also Like

Future of Mail: Drones and Robots Revolutionizing Delivery in 2025

Future of Mail: Drones and Robots Revolutionizing Delivery in 2025 promises faster, smarter, and eco-friendly service—discover how this transformation will change everything.

The Memento Constraint: Why Continual Learning Is the Trillion-Dollar Bottleneck Nobody Is Pricing

Exploring how the inability of current AI models to learn continually shapes the enterprise AI economy and the stakes for future breakthroughs.

Technology operations signal monitor: I admire Fabrice Bellard. He is almost certainly a better overall programmer

A new technology operations signal monitor identifies Fabrice Bellard as an exceptional programmer, emphasizing the importance of role-specific early detection tools.

Inside Apple’s Intelligence Usage Packs For AI

Apple has introduced one-time AI usage packs for Creator Studio. The $49.99 price is reported, but Apple does not specify how much usage a pack buys.