METR’s updated evaluation broadens its time-horizon framework to longer and more open-ended tasks. Its central finding needs two readings: models continue to improve, but selected tasks requiring research judgment remain difficult. In an evaluation-design test, no model-produced task was useful enough for METR to include in its report. That is a limitation on the tested work, not a verdict on every kind of judgment. (metr.org)

The key to interpreting the chart is its unit. A time horizon describes task difficulty using human-reference completion time. It does not measure how long an AI can operate without supervision, and it does not certify that a model can replace an employee for that many hours. (evals.alignment.org)

What a time horizon actually measures

METR’s methodology relates a model’s success on tasks to the time a human would need to complete them. A fitted statistical curve estimates success probability across task durations. The 50% time horizon is where that curve crosses a one-in-two probability of success. It summarizes performance across a task distribution, rather than promising identical reliability on every assignment. (evals.alignment.org)

Consider a hypothetical two-hour horizon. That would mean the fitted curve predicts 50% success on benchmark work associated with two hours of human effort. It would not mean the model works for two hours, succeeds on every shorter task, or can safely be left alone for that period. The threshold and the human-time reference are both essential to the interpretation. (evals.alignment.org)

How to read the trend

The updated chart places model release dates on the horizontal axis and 50% time horizons on a logarithmic vertical axis. Equal vertical distances therefore represent equal ratios, not equal additions of time. Read a point as a benchmark estimate for a model, and its uncertainty interval as part of that estimate—not decorative information around an otherwise exact result. (metr.org)

The report’s accompanying chart data records a combined TH1.1-and-TH2 doubling estimate of approximately 3.6 months, with a 95% confidence interval of 2.08–4.94 months. This verifies that the reported figure appears in METR’s own published results; it is not an independent reproduction of the analysis. It describes growth in the measured horizon, not a doubling of intelligence, output quality or workplace productivity. (metr.org)

That last distinction is consequential. In a January 2026 research note, METR researcher Thomas Kwa warned that doubling a time horizon does not double automation. Even fewer interventions could leave humans handling more complicated failures. Turning benchmark performance into a productivity estimate requires measuring human work around the AI, including prompting, waiting, checking and completing work manually. (metr.org)

Why judgment needs a separate reading

The update distinguishes tasks with readily checkable outcomes from work whose quality is harder to establish. Designing a useful AI evaluation illustrates the latter: producing an artifact is not the same as producing something informative for a research program. METR reports borderline-useful outputs, but none sufficiently useful for inclusion. Progress in a rating category and practical usefulness are different findings. (metr.org)

The appropriate interpretation is narrower than either “AI can automate research” or “AI cannot exercise judgment.” These tests examine selected activities under particular standards. As an analytical matter, a stronger result on a checkable task cannot substitute for evidence on an open-ended one. Conversely, a weak result on one research assignment cannot establish a permanent ceiling on model capability.

Human time is a reference, not a constant

The human side of the comparison also needs scrutiny. METR’s new suite relies substantially on researcher estimates of completion time rather than directly observed, dedicated human attempts. That makes it important to distinguish an estimated reference duration from a stopwatch measurement when interpreting the chart. (metr.org)

There is a broader statistical question, too. In an October 2026 preprint, Drew T. Nguyen and William Fithian reanalyzed earlier METR results using alternative models. They found that human duration did not translate uniformly into AI difficulty: equal proportional increases in human time could represent different increases in difficulty for models. Their proposed estimates performed better under their predictive evaluation criteria. (arxiv.org)

That study predates the November update and should not be presented as an assessment of its new judgment tasks. Its relevant lesson is methodological: the authors recommend reading time horizons alongside diagnostic plots showing how human time relates to AI difficulty. An intuitive unit does not eliminate the need to examine the statistical assumptions beneath it. (arxiv.org)

What would establish workplace readiness?

Kwa’s earlier note identifies several gaps between benchmarks and deployment: isolated versus context-rich work, clearly specified versus ambiguous tasks, and the reliability needed to make delegation worthwhile. He also cautions against treating small differences between model estimates as decisive. These limitations concern how the measurement is used, not whether benchmark progress is real. (metr.org)

For an enterprise evaluation, the practical recommendation is to ask different questions from those answered by the headline curve. Does the test resemble the intended workflow? Who decides whether an answer is acceptable? Can failures be detected before they cause harm? How much human correction is necessary, and does the combined process actually save effort?

A useful deployment trial would set those criteria before comparing systems, then report accepted outcomes and the work needed to obtain them. The time-horizon chart can help frame a capability question. It should not be asked to answer a different question—whether a particular organization can reliably automate a particular job—without evidence from that setting.