Doubling time of the 50%-task-completion time horizon of frontier AI models on METR's software task suite, 2019–2025, in days
- Measures
- The doubling time implied by an ordinary least-squares fit of log(50% task-completion time horizon) against public release date, across the frontier (state-of-the-art at release) AI models METR has evaluated, from GPT-2 (2019) through models released in 2025. The time horizon is the length of task, measured in time taken by skilled human professionals, that a model completes with 50% success on METR's suite of software engineering and AI research tasks (HCAST, RE-Bench, SWAA).
- Population
- Frontier AI models released 2019–2025 that METR has measured on its time-horizon task suite
- Conditions
- METR's task suite and methodology (Time Horizon 1.0, March 2025, 170 tasks; or its 2026 revision TH1.1, 228 tasks, which re-fits the full-period trend using TH1 estimates for pre-2023 models). The value depends modestly on whether the window ends in early 2025 (models through Claude 3.7 Sonnet) or late 2025, and on suite version; both are within the definition and the spread is part of the estimate's uncertainty.
- Unit
- days
Most likely about 203 days; 90% interval 170 to 240 days.
Stated as a distribution, the precision the evidence supports. roughly symmetric around 200 days; spread combines METR's bootstrap uncertainty with the choice of window end and suite version
Assessment
The graph's estimate, all things considered: the figures and arguments it rests on, and how they were combined.
Over 2019–2025, the 50% time horizon of frontier AI models on METR's software and AI-research task suite doubled about every seven months, roughly 200 days, with a 90% range of about 170 to 240 days. The figure comes from METR's fit of log time horizon against model release date. The original March 2025 paper gave 212 days (95% interval 171–249) for models up to Claude 3.7 Sonnet. The revised paper gave 207 days (166–240). Including the 2025 models gives about 196 days (162–223), and the revised 2026 task suite barely moves that (196.5 days). The published figures agree to within about 8%, so the full-period value is well pinned down for this suite and method.
Two caveats limit what the number means. First, the growth rate has not been constant. The evidence supports the view that the time horizon began growing faster around 2024: METR's sub-period fits give about 131–165 days since 2023 and 89–109 days since 2024, and models released in 2025 and 2026 have stayed on the faster line. The full-period figure therefore averages a slower early stretch with a faster recent one, and it overstates the doubling time that applies to current models. How much of the speed-up is real and how much comes from measurement effects is still open. Second, the value is specific to METR's task suite. It depends on how well human baseline times measure task length, and most tasks of eight hours or more use estimated rather than measured times. A bias in those times that grew with task length would change the slope. Whether the trend carries over to real-world software work is a separate question.
The work: which figures, how they were combined, and the arithmetic
Figures, all from METR. (1) Kwa et al., arXiv 2503.14499v1 (March 2025): 212 days, 95% bootstrap CI 171–249, frontier models 2019 to early 2025, 170 tasks. (2) The same paper, v3 (February 2026): 207 days, CI 166–240. (3) METR's "Time Horizon 1.1" post (29 January 2026, metr.org/blog/2026-01-29-time-horizon-1-1/): TH1.0 full-period 195.8 days [162, 223], including 2025 frontier models through Claude Opus 4.5. The hybrid TH1.1 fit gives 196.5 days. Sub-period figures: since 2023, 165.3 [129, 211] under TH1 and 130.8 [107, 161] under TH1.1; since 2024, 108.9 and 88.6 days.
Combination. The definition admits windows ending in early or late 2025, so the median sits between the 196-day and 212-day point estimates, at about 203. The bootstrap intervals are about ±19%, and the published figures sit close together, so the 90% range is set at 170–240. Nothing is added for external validity, because the quantity is defined on METR's suite.
On this pass, the acceleration subclaim was newly assessed as supported (credence 0.78). That confirms the interpretive caveat already in the estimate but does not change the full-period fitted value. A real post-2024 speed-up is already part of the 2025-inclusive fits, and it is why those fits come in lower (196 days) than the early-2025 fit (212 days). The estimate is therefore unchanged.
Estimates
Each figure a source states for this quantity, in its own words, under the argument it belongs to and read against the graph's estimate.
Grey band: the graph's 90% interval, with its central value as a line. Dark marks: each source's own figure. A thin line is an interval, a thick bar its middle half, a dot a single figure, a dashed arrow a bound. An open box is where an argument places the value.
Decomposition
The arguments the estimate rests on, each stated as it runs, with what it shows about the value and its premises linked inline. ↗︎ opens a premise; the map shows how they fit together.
METR measured the 50% time horizon of each frontier model from GPT-2 (2019) onward on tasks whose length is set by human baseline times, and an ordinary least-squares fit of log time horizon against release date gives a doubling time of 212 days (95% bootstrap interval 171–249) for models through early 2025, 207 days (166–240) in the revised paper, and about 196 days (162–223) when 2025 models are included, under both the original and revised task suites. The doubling time is therefore about seven months.
- 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 its premises, the fit pins down the full-period doubling time on METR's suite to within a few weeks: three versions of the analysis give 196 to 212 days. The remaining doubt rests on the reliability of the human baseline times, and only insofar as their errors vary with task length. It also rests on how much the few earliest models pull on the slope. A trend that speeded up recently changes how the figure should be read, but does not change the fitted value itself.
Because the time horizon began growing faster around 2024, with METR's own sub-period fits giving roughly 131–165 days since 2023 and 89–109 days since 2024, a single log-linear fit over 2019–2025 averages a slower early regime with a faster recent one. The full-period figure is then a window-dependent summary: it shifts with the endpoint, with the few early models (GPT-2, GPT-3) that anchor its left end, and with task-suite composition, and it overstates the doubling time that applies to current models.
- 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 supported
Granting that growth sped up around 2024, it follows that the full-period figure averages two regimes and should not be read as the current rate. That premise is now assessed as supported: sub-period fits give about 89 to 109 days from 2024 onward, and the 2025 and 2026 models stayed on the faster line, although measurement-artifact explanations account for some unknown part of the speed-up. The argument does not show that the full-period value is ill-defined. Under a fixed window and suite, the fit is stable to within about 8%, so it bears on how to interpret the figure rather than on what the figure is.
Provenance
Where this quantity's evidence comes from. The figures sources state are under Estimates above.
finding that time horizon has doubled every 212 days with a 95% bootstraped confidence interval 171–249 days
P50 doubling time 195.8 days [162, 223] (TH1.0); 196.5 days (TH1.1)
Furthermore, frontier AI time horizon has been doubling approximately every seven months since 2019, though the trend may have accelerated in 2024.
Overview gives 207 days, 95% CI 166–240 days, for 2019 to early 2025 on METR's 170-task suite.
the length of coding tasks AIs can handle, their “time horizon”, doubled every 7 months from 2019 - 2024 and every 4 months from 2024-onward.
Citing a METR report on AI task time horizons.
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.
Created by extractor · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.