The “FiveHour AI” Myth—and What It Really Says About Your Job -The chart everyone shares (and misreads)

A nonprofit called METR (Model Evaluation & Threat Research) became famous for a graph that tracks how quickly frontier AI models are improving at completing longer tasks, with results that look roughly exponential over recent years. METR’s own write-up describes a “doubling time of around 7 months” for the length of tasks models can complete (measured in human time), and notes that models are near-perfect on tasks that take humans under a few minutes but fall below 10% success on tasks taking more than ~4 hours (in their suite).​ 

That framing matters because the y-axis is not “how long the model can work nonstop” or “how many hours of a job it can replace.” It’s a time horizon: the human time-to-complete of tasks at which the model hits a chosen success probability (often 50%), based on a fitted curve relating task length to model success.​ 

What “time horizon” actually measures 

METR builds the metric by collecting software-engineering-relevant tasks, timing human experts, and then testing models to see how success drops off as tasks get longer for humans; the “time horizon” is essentially the point on that human-time scale where the model’s predicted success rate crosses a threshold like 50%. This is why it’s easy to over-interpret a datapoint like “~5 hours” as “it can do five hours of autonomous labor,” when it really means “it sometimes succeeds on tasks that take humans about five hours”—and “sometimes” can include wide uncertainty depending on task choice, evaluation design, and model variance.​ 

Even METR emphasizes a broader contradiction the public feels: models can look superhuman on tightly-scored knowledge tests, yet still fail at end-to-end work projects, because the hard part is often stitching many steps together reliably, not answering a single question.​ 

A reality check from real work, not benchmarks 

One reason the jobs debate gets overheated is that benchmark performance does not translate cleanly into workplace productivity. METR ran a randomized controlled trial with 16 experienced open-source developers working on real issues in codebases they knew well (246 issues total), and found that when developers were allowed to use AI tools (typically Cursor Pro with Claude Sonnet-class models at the time), they took 19% longer to complete issues on average. 

The most uncomfortable detail is the perception gap: developers expected AI to speed them up by 24%, and even after the experiment they still believed AI had sped them up by 20%, despite the measured slowdown. METR’s takeaway isn’t “AI is useless,” but “translation is hard”: realism introduces friction—implicit requirements, quality bars, context, and iteration—that many benchmarks systematically minimize. 

So…doom for jobs, or a wave of opportunity? 

The best evidence points to a more nuanced story: disruption is real, but “tasks” are more automatable than “jobs,” and the net effect depends on adoption speed, complementary investments, and who captures the gains. The World Bank’s World Development Report on the changing nature of work argues that routine, codifiable tasks are most vulnerable, while technology also creates new tasks and sectors—and that investing in human capital and social protection is central to navigating the transition.​

In other words, the near-term risk is less “everyone gets replaced overnight” and more “some work gets unbundled, re-priced, and re-assigned,” with faster change in occupations heavy on routine information processing and slower change where context, accountability, and human interaction dominate. That same World Bank report explicitly warns against sensational predictions while still acknowledging that automation can eliminate many low-skill jobs in some contexts, even as it opens new opportunities elsewhere.​ 

What to conclude from the METR graph without getting fooled 

METR’s plot is valuable as a capability trend signal in a particular evaluation regime, mainly centered on software-style tasks and measured against human time, and it’s plausible that it continues to move fast. But it is not a stopwatch for “how long an AI employee can run,” nor a direct forecast of mass unemployment on its own—especially when METR’s own realistic study found slowdown rather than speedup for experienced developers in a high-context setting.​ 

If you want one grounded way to say it: AI is getting better quickly at certain long, structured tasks, yet the labor-market impact will be mediated by workflow redesign, verification costs, organizational risk tolerance, and skills adaptation—so the outcome is neither clean doom nor guaranteed boom.​ 

Quoted directly from METR’s own description of their work: “We propose measuring AI performance in terms of the length of tasks AI agents can complete,” and they caution that “the best current models… are capable of some tasks that take even expert humans hours, but can only reliably complete tasks of up to a few minutes long.”