---
title: "Every 3.6 Months, the Tasks Double. Every Year, That's Ten Times."
date: 2026-11-22
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-metr-doubling-time-1122
language: en
---

# Every 3.6 Months, the Tasks Double. Every Year, That's Ten Times.

> Twelve divided by 3.61 is 3.32. That is the arithmetic behind the headline number in METR's November capability update, published on November 20.

Twelve divided by 3.61 is 3.32.

That is the arithmetic behind the headline number in METR's November capability update, published on November 20. The evaluation organization measures how long a task a model can finish on its own: a goal, no hints, no human picking up the work halfway. On the task sets it calls TH1.1 and TH2, that duration has been doubling about every 3.61 months.

Twelve months is 3.32 doublings. Two to the power of 3.32 is about 10.

## What is actually being measured

METR does not sell models and does not run a leaderboard. That shapes how the number reads.

The tasks have defined starts and ends, and the outcomes are checkable. A model either finishes or it does not. A duration is harder to game than a benchmark percentage.

The confidence interval around 3.61 months runs from 2.08 to 4.94. That range is wide enough to hold two very different stories about the next two years. Measured earlier on TH1.1 alone, the figure was 105 days. Adding TH2 moved the estimate to 3.61 months, about 110 days. The direction held; the speed shifted slightly. The interval did not narrow. The measurement is still moving.

## The other division: 154x

METR also carved out TH2 as a verifiable subset, where scoring is automatic.

On that subset the doubling period is 1.65 months. Twelve divided by 1.65 is 7.27 doublings. Two to the power of 7.27 is about 154.

It is not the same number as 10.

The 10x comes from TH1.1 and TH2 together. The 154x comes from TH2 alone, the part with unambiguous right answers that a script can grade. A smaller subset, with more regular tasks, moves faster. Set the two side by side and it is easy to read them as one thing that grew 154-fold. The difference is the denominator: one is a union of two task sets, the other is a slice of it.

The slice is also where models improve fastest. Code that runs against tests either passes or fails. Math has an answer. A bug fix either turns the test green or it does not. None of that means every kind of work compounds at the same rate. Tasks without a score, such as deciding which clause in a contract is the problem, cannot enter this curve at all, because there is nothing to grade.

So: about 10x a year across the combined set, about 154x a year on the verifiable subset. Anyone who reports 154x as capability growing 154-fold has swapped the subset for the whole.

## Cost enters the measurement

This update adds a dimension METR calls lifecycle cost.

Some tasks run in a clean container. Others require building an environment, installing dependencies and running the scoring harness 100 times. The expensive subset costs 2.9 times more than the cheap one, and a model's time horizon on it is 2.9 times worse, which equals about 1.5 doublings. Two to the power of 1.5 is 2.83, close enough to 2.9 that the conversion holds together.

Then the sharper result. For tasks of equal length, each doubling of cost pushes the point at which a model reaches 50% success about 4.6 months further out.

4.6 months is longer than one doubling period. The model is improving while the environment gets more expensive, and the second is eating part of the first. Hold the model fixed and give it a costlier environment, and its effective horizon moves backward by more than a full doubling.

The correlation between task duration and lifecycle cost has an R² of 0.82. The two curves nearly lie on top of each other.

## The model grading itself

METR also tested judgment, and the result is uncomfortable.

Asked to review output from itself or a stronger model, a model systematically overrates quality. Claude Opus 5.5 scored Astra tasks about 0.2 points higher than human raters, on a 1-point scale.

0.2 sounds small. It is not noise. Noise cancels; bias does not. A consistent direction propagates, and if those scores feed training, the next round learns the bias.

On the LMCA benchmark, every public model fell below the consistency of an estimate a domain expert produces in 15 minutes. A model that can work unattended for 200 hours cannot reliably judge a document a person reads in a quarter of an hour. Those are different abilities: one is whether it can do the work, the other is whether it knows how well it did.

## July 2027, with a caveat attached

Extrapolate 3.61 months forward without adjustments, and by July 2027 a model has a 50% chance of completing the longest scored threshold in the task set: a 200-hour task.

The 50% is the whole sentence. Not a promise, but even odds. Extrapolation is not a forecast either; it is the current slope extended into next summer. Slopes bend, task sets get replaced, measurement methods change. METR is stating a conditional: if the line holds, it is a coin flip by then.

200 hours is about five working weeks. One task, five weeks, no handoff.

## Two divisions, one denominator

Go back to the arithmetic.

3.32 doublings, 10x. 7.27 doublings, 154x. Both correct, different denominators.

Past capability curves divided by parameters, tokens or training compute. This one adds a line item: what it costs to finish the task. That is the detail worth keeping, because unlike the doubling rate, cost is something a team can change. A doubling of cost is worth 4.6 months, and that lever sits on the engineer's side of the table.

The 10x is rising. So is the denominator.

The verifiable subset says 154x. The combined set says 10x. Neither is an endpoint. They are two long divisions, and the one that will decide what gets deployed is the one with a dollar sign in it.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-metr-doubling-time-1122
Free to quote with attribution and a link to the original.
