AI at Work Observatory

How model capabilities are evolving.

What length of task can an AI model complete on its own, and how fast is that length growing?

The evaluation organisation METR measures the time an expert human needs for the tasks a model completes on its own half the time. Its update of 20 November 2026 covers fifteen public models from OpenAI and Anthropic.

From 7 minutes to 40 hours.

Three figures and a curve: where the models were, where they are, and how fast they are progressing.

Yesterday · May 2024 model

7 minutes

GPT-4o, the oldest of the models tested, completed tasks of this length half the time.

Today · September 2026 models

≈ 40 hours

The length measured for GPT-6 Astra and Claude Opus 5.5: about 340 times that of May 2024. The uncertainty is great: from 13 to 2,645 hours for GPT-6 Astra.

Trend

×2 in 3.6 months

Doubling rate measured from May 2024 to September 2026. Models released since then and the laboratories’ internal models are not measured.

Length of the tasks completed half the time, by model release date

Logarithmic scale. Vertical lines: 80% confidence interval.

After METR, November 2026 update (CC BY): data and findings from METR, adapted and translated. Figures taken on 22 November 2026.

See the values by model
50% time horizon on the TH1.1 and TH2 task suites combined (METR, November 2026)
Model and release date50% horizon95% interval
GPT-4oMay 20247 min4 to 14 min
o1December 202435 min16 to 66 min
GPT-5August 20253 h1.6 to 5.6 h
Claude Opus 4.5November 20255 h2.4 to 8.5 h
Claude Opus 4.6February 20266 h3.2 to 12 h
GPT-5.4March 20266 h3.3 to 9.9 h
Claude Opus 4.7April 20268 h4.3 to 16 h
GPT-5.5April 202611 h6.1 to 19 h
Claude Opus 4.8May 202612 h6.1 to 21 h
Claude Fable 5June 202616 h8.9 to 32 h
GPT-5.6 SolJuly 202612 h6.6 to 21 h
Claude Opus 5July 202620 h9.4 to 86 h
Claude Fable 5.1September 202630 h12 to 308 h
GPT-6 AstraSeptember 202640 h13 to 2,645 h
Claude Opus 5.5September 202639 h12 to 1,253 h

METR, report of 20 November 2026 · METR, definition of the time horizon

Strengths and weaknesses of current models.

Strong at coding.

On programming and optimisation tasks, models usually match or outperform skilled humans.

Weaker on costlier tasks.

A time horizon 2.9 times shorter on the tasks that would cost most to turn into training environments.

Weaker at judging.

Without concrete feedback on the result, their judgement remains worse than that of humans.

What the measure does not say.

This measure is about tasks, not occupations. Three points to keep in mind.

A human duration.

The time horizon is the time an expert human needs for the task, not the time the model works on its own. A model is usually faster on the tasks it completes.

Mostly software tasks.

Programming, optimisation, machine learning, games and a few research tasks: the suite does not cover the variety of occupations. It measures tasks, not work as it is actually done.

An increasingly fragile measure.

Models saturated the first task suite within a year. For the new one, human time is usually estimated, not measured, and the confidence intervals of the latest models are very wide.