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

🔍 Read the full analysis: The Hours Behind AI Capability: What Teams Should Understand on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

METR’s public benchmark records an 80% success time horizon of 185.9 minutes for an early Claude Mythos Preview released in April 2026. The measure estimates the human working time required for tasks an AI agent can complete reliably; it does not mean the AI works at human speed or succeeds on every task. METR’s public record has not been updated since September 2026.

A November 18 analysis of METR’s public benchmark results puts the best measured AI-agent 80% task horizon at 185.9 minutes—about three hours—based on an early Claude Mythos Preview released in April 2026. The estimate measures the length of tasks, in skilled human working time, that an agent completes at a given success rate; it is a planning yardstick, not a promise that a model will finish every three-hour job.

METR, an independent research nonprofit, estimates a time horizon by timing skilled professionals on tasks, asking AI agents to attempt those tasks, then fitting a curve showing how success changes with task length. The 80% figure is the point where the agent is predicted to succeed on four out of five attempts. METR’s Time Horizon 1.1 suite contains 228 tasks spanning software engineering, machine learning and cybersecurity. The evaluation uses six independent runs per task and treats cheating detected in transcripts as failure.

The source analysis reports that the April 7, 2026 Claude Mythos Preview result is 185.9 minutes at 80% reliability. Its confidence interval is wide, from 97.3 to 398.5 minutes, or roughly 1.6 to 6.6 hours. The analysis also lists prior public results: 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 figures are benchmark results, not a guarantee of performance on a company’s own work.

Reliability changes the estimate sharply. For the same Mythos Preview, the reported 50% horizon is 1,044.8 minutes, or about 17.4 hours, compared with 3.1 hours at 80%. METR cautions that measurements above 16 hours are unreliable with its current task suite. The source says the public results page was last updated September 8, 2026, and is no longer actively updated; the figures do not establish the state of models released afterward.

At a glance
reportWhen: Analysis published November 18, 2026; t…
The developmentA November 18, 2026 analysis of METR’s benchmark results says the best publicly measured 80% AI task horizon was about three hours, while warning that benchmark conditions and reliability limit what teams can infer.

Turning Benchmarks Into Work Plans

For team leads, the measure offers a more practical question than whether AI “can” do a task: how long would a skilled person take, and how often does the agent complete comparable work? That framing can help teams divide projects into small, testable units, choose where to try automation, and estimate how much review time to reserve.

The result also argues against handing over a long project as one unmonitored assignment. At an 80% success threshold, one in five attempts is still expected to fail under the benchmark’s fitted estimate. The source analysis describes this as a reason to build in human checking, clear acceptance tests and escalation paths, rather than treating a benchmark horizon as a capacity guarantee.

Real work may be less structured than benchmark tasks. The article cites a METR comparison in which messier tasks reduced the measured horizon by almost a factor of three. It also reports a 2025 METR trial where experienced developers took 19% longer with AI while believing they were 20% faster, and a separate study in which reviewers caught 31% of errors requiring task-specific rules to identify. Those findings, as presented by the source, show why local evaluation and review matter; they do not establish that every team or tool will have the same results.

Philips 5500 Series Fully Automatic Espresso Machine with LatteGo Milk System

Philips 5500 Series Fully Automatic Espresso Machine with LatteGo Milk System

KitchenAid Classic Series 4.5-Quart Tilt-Head Stand Mixer, Onyx Black

KitchenAid Classic Series 4.5-Quart Tilt-Head Stand Mixer, Onyx Black

Shark Rocket Pet Plus Corded Stick Vacuum, Bordeaux/Silver, HV322

Shark Rocket Pet Plus Corded Stick Vacuum, Bordeaux/Silver, HV322

As an affiliate, we earn on qualifying purchases.

How METR Measures Task Length

The horizon is based on human completion time, not how long the AI itself spends working. METR describes it as the duration of a task, measured by expert completion time, at which an AI agent is predicted to succeed with a specified reliability. According to the source, the professionals recruited for the benchmark average about five years of relevant experience.

Short tasks generally produce higher success rates than long tasks, so METR fits a declining success curve across task lengths. The point where that curve crosses 80% is the 80% horizon; the 50% crossing gives a less stringent measure. METR says agents that do complete benchmark tasks are typically several times faster than humans, but that speed is distinct from the human-time scale used to describe task difficulty.

The source analysis says the public 80% record rose from less than a second for GPT-2 in 2019 to 0.89 minutes for GPT-4 in March 2023 and 66 minutes for GPT-5.2 in December 2025, before the reported 185.9-minute result in April 2026. METR’s February-to-March 2026 Frontier Risk Report estimated the labs’ internal frontier at “≥3h but <4h” at 80%, according to the article. That estimate is separate from the public benchmark record.

Limits of the Public Estimate

The reported record has a broad 97.3-to-398.5-minute confidence interval, and the benchmark’s clean, well-specified tasks may not match everyday company work. The source says messier tasks cut the horizon by almost a factor of three in one METR comparison, but it does not provide enough detail here to determine how that result applies across different roles or task types.

METR’s public results page was last updated September 8, 2026, and is described as no longer actively updated. The analysis supplies no public measurement after the April 2026 Mythos Preview result. It is therefore unclear from this material how the newest systems perform on the same evaluation, whether the benchmark suite or scoring has since changed, or how reliably any particular model handles a company’s own backlog.

The 80% figure is an estimated success rate across tasks of a given length, not a per-task guarantee. The 50% estimate is even less suited to routine planning, and METR’s stated reliability warning applies above 16 hours. The article’s projected dates for an eight-hour 80% horizon depend on trend fits and are forecasts, not measured outcomes.

Testing Against Team Backlogs

The source recommends that teams start with work divided into hour-scale pieces with clear acceptance checks, then assign AI the parts that can be tested objectively. Managers should budget time for review, select models according to the task and measure performance on their own work every quarter.

Those local tests can show whether benchmark results transfer to a team’s code, processes and error costs. The next public comparison would require updated measurements using a clearly described task suite and reliability threshold. The source does not identify a date for a new METR update, so teams should treat the cited three-hour figure as the latest public result covered by the analysis, not a current measurement of every model.

Key Questions

What does a three-hour AI time horizon mean?

It means the benchmark estimates that an agent can complete tasks taking a skilled person about three hours with an 80% success rate. It does not mean the agent works for exactly three hours or will finish every such task.

Which model had the highest result in the cited public data?

The source analysis identifies an early Claude Mythos Preview, dated April 7, 2026, as the public record-holder in METR’s results file, with a 185.9-minute 80% horizon. The confidence interval reported for that estimate is 97.3 to 398.5 minutes.

Why is the 50% time horizon much longer?

A 50% threshold requires success only half the time, while the 80% measure requires four successes out of five. The source reports 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 teams use the benchmark to forecast project completion?

Not on its own. The benchmark can inform task selection, but the source says real-world tasks may be messier and recommends testing on a team’s own work, setting acceptance checks and budgeting for human review.

Source: ThorstenMeyerAI.com

You May Also Like

Wordgard: In-browser Rich-text Editor From The Creator Of ProseMirror

The creator of ProseMirror introduces Wordgard, a new in-browser rich-text editor aimed at developers and content creators, now available for testing.

Your Coding Agent Is an Attack Surface: The Claude Code Security Reckoning

Recent vulnerabilities in Claude Code reveal critical attack surfaces in AI developer agents, risking token theft and code execution.

Local AI Trends: Muse Glimmer Emerges

An IdeaNavigator AI brief flags Muse Glimmer as a monitoring signal, but provides no independent details confirming the model or its capabilities.

Putting Linux On My RTX Spark Laptop Because I Don’t Have A Microsoft Account

Developer Josephine Pfeiffer reports booting Arch Linux ARM on an ASUS ProArt P16 with NVIDIA’s N1X chip and accessing a 116 GiB memory pool.