Strong at coding.
On programming and optimisation tasks, models usually match or outperform skilled humans.
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.
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
After METR, November 2026 update (CC BY): data and findings from METR, adapted and translated. Figures taken on 22 November 2026.
| Model and release date | 50% horizon | 95% interval |
|---|---|---|
| GPT-4oMay 2024 | 7 min | 4 to 14 min |
| o1December 2024 | 35 min | 16 to 66 min |
| GPT-5August 2025 | 3 h | 1.6 to 5.6 h |
| Claude Opus 4.5November 2025 | 5 h | 2.4 to 8.5 h |
| Claude Opus 4.6February 2026 | 6 h | 3.2 to 12 h |
| GPT-5.4March 2026 | 6 h | 3.3 to 9.9 h |
| Claude Opus 4.7April 2026 | 8 h | 4.3 to 16 h |
| GPT-5.5April 2026 | 11 h | 6.1 to 19 h |
| Claude Opus 4.8May 2026 | 12 h | 6.1 to 21 h |
| Claude Fable 5June 2026 | 16 h | 8.9 to 32 h |
| GPT-5.6 SolJuly 2026 | 12 h | 6.6 to 21 h |
| Claude Opus 5July 2026 | 20 h | 9.4 to 86 h |
| Claude Fable 5.1September 2026 | 30 h | 12 to 308 h |
| GPT-6 AstraSeptember 2026 | 40 h | 13 to 2,645 h |
| Claude Opus 5.5September 2026 | 39 h | 12 to 1,253 h |
METR, report of 20 November 2026 · METR, definition of the time horizon
On programming and optimisation tasks, models usually match or outperform skilled humans.
A time horizon 2.9 times shorter on the tasks that would cost most to turn into training environments.
Without concrete feedback on the result, their judgement remains worse than that of humans.
This measure is about tasks, not occupations. Three points to keep in mind.
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.
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.
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.