Doubling time, in days, of the 50%-task-completion horizon of frontier AI models released from 2024 onward on METR's software task suite
- Measures
- Doubling time implied by a least-squares fit of log(50% time horizon) against release date, over frontier (state-of-the-art at release) models released from 1 January 2024 onward, on METR's software and AI-research task suite
- Population
- Frontier AI models released 2024 onward evaluated by METR
- Conditions
- METR time-horizon methodology, TH1.0 or TH1.1 task suite; window end as of latest published METR fit
- Unit
- days
Most likely about 103 days; 90% interval 82 to 132 days.
Stated as a distribution, the precision the evidence supports. Centred on METR's latest published fit (105 days), right-skewed toward longer values from fit sensitivity and saturation
Assessment
The graph's estimate, all things considered: the figures and arguments it rests on, and how they were combined.
METR's published fits of the 50% time horizon of frontier models released from 2024 onward put the doubling time at roughly three and a half months. The original task suite gave 109 days and the revised Time Horizon 1.1 suite gave 89 days (January 2026). The latest published METR fit, in its May 2026 Frontier Risk Report, covers public frontier models through early 2026 and gives 105 days with an R² of 0.98. An independent regression on METR's published model file finds about 3.5 months, with a 95% interval of 3.2 to 3.8 months. The best estimate is therefore about 103 days, with a 90% interval of roughly 82 to 132 days. That is about 10 to 12 doublings' worth of growth per year, or around tenfold per year, against a seven-month doubling over 2019 to 2025.
The interval is wider than any single fit's statistical interval for two reasons. First, the fitted value moves by 15 to 20 days depending on the task suite version and which recent models are included. Second, the newest frontier models measure 12 to 17 hours, and METR says its suite cannot reliably measure horizons above about 16 hours. The recent end of the fit is therefore poorly pinned, and a refit on a new suite with longer tasks could move the figure either way. The figure also rests on growth being roughly exponential over the window and on METR's human baseline times, most of which are estimated rather than measured for the longest tasks.
The work: which figures, how they were combined, and the arithmetic
Figures: TH1.0 since-2024 fit 109 days and TH1.1 89 days (METR, 2026-01-29, metr.org/blog/2026-01-29-time-horizon-1-1/). Fit to public frontier models released after 1 January 2024, 105 days, R² 0.98 (METR Frontier Risk Report, 2026-05-19, Appendix E, metr.org/blog/2026-05-19-frontier-risk-report/). Independent regression over the 21 models since 2024 in METR's file: 3.5 months, 95% CI 3.2 to 3.8 months, about 97 to 116 days (iqmetrics.org, October 2026). For comparison, METR's 2023-onward TH1.1 fit is about 129 days (95% CI 104 to 158) and the all-time fit is about 188 days. Opus 4.6 was revised from 14.5 to about 12 hours after a regularization correction in March 2026.
Combination: the definition takes the latest published METR fit, which is 105 days. The median is placed slightly below that, at 103 days. The independent reanalysis agrees with 105 days, the TH1.1 figure was 89 days, and saturation compresses the top measurements, which biases the fit toward longer doubling times. The statistical interval of the fit (about plus or minus 10%) is widened to cover the 89 to 109 day spread across suite versions and model sets. It is widened further on the high side because a future refit with saturated or re-baselined long tasks could lengthen the figure. Conversion: 365/103 gives 3.5 doublings per year, a factor of about 11.7 per year.
No newer METR fit than May 2026 was found in searches dated to November 2026.
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.
The fitted doubling time depends heavily on the newest frontier points, and horizons above roughly 16 human-hours are unreliable on the current suite; the top models of early 2026 sit at 12 to 17 hours with confidence intervals spanning several-fold. The recent end of the fit is therefore poorly pinned down, and the compression of measured horizons near the ceiling tends to lengthen the fitted doubling time relative to underlying capability growth.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting that horizons above about 16 hours are unreliable on the current suite, the newest points are noisy and compressed. That widens the plausible range of the fitted value and tilts underlying growth toward being faster than fitted. Only the last two or three models are near the ceiling, so the effect on a fit over about twenty models is modest. The argument mainly justifies a wider interval rather than a different central value.
METR's least-squares fits of log 50% horizon against release date over frontier models released since 1 January 2024 gave 109 days under TH1.0 and 89 days under TH1.1 (January 2026), and 105 days with R² of 0.98 in the May 2026 Frontier Risk Report, which adds models through early 2026; an independent regression on METR's published model file gives about 3.5 months. Given that growth over the window is approximately exponential and that the human baseline times reliably measure task length, the value is about 90 to 110 days, with 105 days the latest published METR figure.
- 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?
- 2 not separately assessed
Granting its premises, the published fits pin the value to roughly 90 to 110 days. They are the quantity's own defining computation, and an independent regression on METR's file reproduces them. The argument leans most on approximately exponential growth over the window, which the high R² supports, and on reliable human baseline times, which is weaker for the long tasks whose times are estimated. Neither premise has yet been assessed on its own.
Provenance
Where this quantity's evidence comes from. The figures sources state are under Estimates above.
We also report below the doubling time since 2024: this was at 109 days under TH1, and falls to 89 days under TH1.1.
This fit has a doubling time of 105 days, and an R2 of 0.98.
the 21 models released since 2024 give 3.5 months (95% CI 3.2 to 3.8), close to the 105 days in METR's own risk report
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 claim_steward · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.