🔍 Read the full analysis: AI And The Clock: What Task Duration Reveals About Capability on ThorstenMeyerAI.com
Get the latest gadgets 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
METR’s public benchmark records an 80% task horizon of 185.9 minutes for an early Claude Mythos Preview, based on tasks timed for skilled human workers. The measure describes task difficulty, not how long the AI takes to work, and its results do not guarantee success on messier workplace assignments. METR’s public results have not been actively updated since September 2026.
A benchmark record cited in a November 18 analysis puts an early version of Anthropic’s Claude Mythos Preview at an 80% task horizon of 185.9 minutes on METR’s evaluation, meaning it succeeded on tasks that skilled people typically took about three hours to complete in four out of five trials. The result offers a measure of AI capability that is more specific than asking whether a model can perform a type of work, but it is not a guarantee that the model can reliably handle comparable tasks in a company’s own systems.
METR, an independent research nonprofit, defines a time horizon as the length of a task, measured by skilled human completion time, at which an AI agent is predicted to succeed at a specified reliability level. The measure refers to how difficult the task is—not how long the AI itself spends on it. METR says agents can complete successful tasks several times faster than people.
The analysis reports that METR’s Time Horizon 1.1 suite includes 228 tasks, timed with skilled professionals in fields including software engineering, machine learning and cybersecurity. METR conducts six independent runs per task and treats cheating detected in transcripts as failure. It then estimates where success rates fall below thresholds such as 80% or 50% as task duration increases.
In the results file cited by the analysis, 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 early Claude Mythos Preview result is dated April 7, 2026. Its reported confidence interval is wide: 97.3 to 398.5 minutes. METR’s public results page says it is no longer actively updated and was last updated September 8, 2026.
Turning Benchmarks Into Work Plans
The time horizon can help teams frame AI work around task length and reliability rather than a model’s most impressive demonstration. For a team deciding what to delegate, a result near three hours at 80% success suggests a tested capability under benchmark conditions—not that any three-hour assignment will be completed correctly.
That distinction matters for budgeting. An 80% success rate means one failure in five attempts on the relevant evaluation, before accounting for differences between benchmark tasks and a company’s code, data, permissions or review process. Teams still need acceptance tests, human review and time to repair errors. The analysis argues for dividing work into clear, testable pieces and measuring performance on the team’s own tasks.
The source also reports that a separate METR comparison found messier tasks reduced the horizon by almost a factor of three. That is a warning against directly translating a clean benchmark result into a staffing forecast. The number may help compare capabilities over time, but it does not by itself establish productivity gains or return on investment for a particular organization.
Products Worth Considering

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

Jackery HomePower 3000 Portable Power Station, 3072Wh

Samsung Galaxy Tab A11+ Kids Edition, 128GB, Royal Blue
As an affiliate, we earn on qualifying purchases.
How METR Measures Task Length
METR’s approach ties model performance to the time skilled people need for a task. Human professionals complete and time the benchmark tasks; AI agents then attempt those tasks repeatedly. The resulting success rates are plotted against human task duration, and the fitted curve is used to estimate a horizon at a chosen reliability level. The 80% threshold is the stricter planning figure discussed in the analysis; the 50% horizon marks a much less reliable level.
The two thresholds can produce very different impressions. The early Mythos Preview’s reported 50% horizon is 1,044.8 minutes, or about 17.4 hours, compared with 185.9 minutes at 80%. METR cautions that measurements above 16 hours are unreliable with its current task suite, so the longer estimate should be treated carefully. A horizon is also an estimated average across tasks, not a promise that every task of that length will have the same outcome.
The analysis cites METR’s February-to-March 2026 Frontier Risk Report, which estimated the labs’ internal frontier at at least three but under four hours at 80%. That estimate provides context for the public result, but it is not a continuing public measurement of current models. The source’s discussion of possible future horizons is a forecast based on past trends, not a confirmed capability milestone.
Limits of the Three-Hour Estimate
The cited result is an estimate from a defined benchmark, not a direct test of every workplace setting. Its confidence interval spans about 1.6 to 6.6 hours, and METR says its task suite is less reliable for measurements above 16 hours. The analysis also reports a sizable reduction when tasks are messier, though it does not provide enough detail here to generalize that factor to every job or model.
It is not clear from the supplied material whether METR has released newer public measurements since the page’s September 8 update. The source says nothing after April 2026 has been measured publicly. Nor does the benchmark establish how often a company’s own AI workflows will fail, how much supervision they require, or whether using them saves time overall.
The source cites a 2025 trial in which experienced developers were 19% slower 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 are reported as context in the analysis; the supplied material does not provide study details sufficient to apply their figures to all teams or current models.
Testing Models on Real Work
The practical next step, according to the source analysis, is for teams to test AI on their own tasks rather than rely only on a general benchmark. That means choosing work with clear acceptance checks, recording completion and error rates, and including the time spent reviewing and correcting results.
The analysis recommends breaking work into pieces measured in hours, starting with tasks that can be tested, and matching models to the job. It also proposes re-measuring performance on internal work every quarter. These are recommendations from the source, not a formal METR requirement.
For the broader capability trend, the key development to watch is whether METR resumes or expands public evaluations and reports results on less controlled tasks. Until then, the April 2026 record and the benchmark’s stated limits are the confirmed basis in the supplied material; a full workday at 80% remains a forecast, not an established public measurement.
Key Questions
What does an 80% AI time horizon mean?
It is the estimated human task duration at which an AI agent succeeds about four times out of five under the benchmark’s conditions. It does not mean the AI takes that long to complete the task.
Which model had the highest result cited?
The source analysis cites an early Claude Mythos Preview, dated April 7, 2026 in METR’s results file, with an 80% horizon of 185.9 minutes. Its reported confidence interval is broad, from 97.3 to 398.5 minutes.
Does the result mean an AI can reliably do any three-hour job?
No. It describes performance on METR’s benchmark tasks, not every workplace assignment. The source reports that messier tasks reduced the horizon in one comparison, and teams must account for their own work, testing and review.
Why is the 50% horizon much longer?
A model can reach a longer task duration when success is required only half the time. The source gives 1,044.8 minutes at 50% for the early Mythos Preview, compared with 185.9 minutes at 80%, while noting that measurements above 16 hours are unreliable with the current task suite.
Are newer public results available?
The supplied analysis says METR’s public results page was last updated September 8, 2026 and is no longer actively updated. It reports no public measurements after April 2026, so later capability changes are not established by the cited material.
Source: ThorstenMeyerAI.com
