The task-completion time horizon of frontier AI models began growing faster around 2024.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
On METR's software task suite, the 50% time horizon of frontier models has grown roughly twice as fast since about 2024 as over the full 2019–2025 record. The full-period doubling time is about 200 days. The post-2024 doubling time is about 103 days, with a 90% interval of roughly 82–132 days. METR's latest published fit gives 105 days. Models released after the fit, through Claude Opus 4.6 at about 12 hours, stayed on the faster line.
The case falls short of verified for three reasons. The evidence comes from one organization's suite and about a dozen models. The rate shifted by about a fifth when the suite was revised. And the long tasks that recent models are scored on mostly have estimated rather than measured human times. A reanalysis argues that the human-time scale may make the 2–30 minute band unusually easy to cross. That could inflate progress in 2023–24, but it does not obviously explain the growth above an hour in 2025–26. The claim that measurement artifacts explain the acceleration is therefore possible but looks unlikely. Two kinds of evidence would settle the question: a suite re-baselined with measured long-task times, or independent benchmarks that test for a break.
Full reasoning: the evidence and decisions behind this verdict
This pass was prompted by the post-2024 doubling-time quantity receiving its first estimate: median 103 days, 90% interval 82–132. The quantity's Steward cites METR's Frontier Risk Report (May 2026), which gives 105 days with R² 0.98. This agrees with the earlier figures (88.6–108.9 days since 2024 in the Time Horizon 1.1 post, metr.org/blog/2026-01-29-time-horizon-1-1/). Set against the full-period estimate (median 203 days, 90% interval 170–240), the post-2024 interval sits entirely below the full-period interval, and the ratio is about 0.5.
Correction: the previous assessment cited a Claude Opus 4.6 horizon of about 14.5 hours. METR corrected this to about 12 hours in March 2026, as noted by the quantity's Steward; this pass did not check it independently. Recomputing with the corrected figure: Claude 3.7 Sonnet was at about 1 hour (February 2025) and Opus 4.6 at about 12 hours (February 2026). That is log2(12) ≈ 3.6 doublings in about 365 days, or roughly 100 days per doubling. This still matches the faster rate and is no longer faster than it. The conclusion does not change.
The counter-evidence is unchanged since the earlier pass: - Nguyen & Fithian, arXiv 2610.12466, find a flat difficulty mapping between 2 and 30 minutes. - The task-suite revision shifted the rate by about 20%. - The long-task human times are mostly estimates. - The suite is near saturation.
Each of these plausibly explains part of the speed-up, not all of it. The status stays supported, with credence 0.78.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
METR's log-linear fit over all frontier models gives a full-period doubling time of about 200 days. The same fit restricted to models released from 2024 onward gives a post-2024 doubling time of about 90–110 days, roughly half. Models released later in 2025–26 (Claude Opus 4.5 at about 5 hours, Opus 4.6 at about 12 hours after METR's March 2026 correction) stayed on the faster line. Task length is measured in human baseline times, taken as a consistent scale, so the horizon grew faster from about 2024.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 3 not separately assessed
Granting its premises, the conclusion is much more likely than not. Those premises are a post-2024 doubling time of about 103 days, against a full-period figure of about 200, and later models that stay on the faster line. Together they make a faster growth rate from around 2024 much more likely. It falls short of certainty because the recent fit rests on about a dozen models and the break date is not pinned down. It also leans on human baseline times being a consistent scale, which is weakest for long tasks.
The post-2024 speed-up rests on about ten frontier models with wide confidence intervals, changed by about a fifth when METR revised its task suite, uses estimated rather than measured human times for most tasks of eight hours or more, and may reflect a human-time scale on which crossing roughly 2–30 minutes is unusually easy; if these measurement artifacts explain the apparent acceleration, horizon growth did not really speed up.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting that measurement artifacts explain the apparent acceleration, it follows that horizon growth did not really speed up, so the argument stands or falls on that premise. Each listed artifact is real. Each plausibly accounts for only part of the speed-up, and none obviously explains the growth above one hour in 2025–26. On current evidence the premise looks more likely false than true.
The claims this one rests on directly, not gathered into a named line of reasoning.
- supportsthis provides evidence for the parentsteward instructions →Benchmarks beyond METR's software suite show AI time horizons or capabilities growing faster since 2024. ↗︎
Provenance
Where this claim has been said, linked to its canonical form.
though the trend may have accelerated in 2024
Paper summary reports an exploratory 2023–2025 fit with a doubling time of about 172 days.
Now the update confirms: the trend from 2024 on looks like it’s doubling much faster than once per 7 months.
Commentary on METR's Time Horizon 1.1 release, arguing growth went from about 3x/year before 2024 to about 10x/year.
Assessment history
0 status changes over 2 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.