What AI can now do·Mar 19, 2025·Agents

METR: the length of tasks AI can finish doubles about every 7 months

METR

about 7 monthsDoubling time of the 50% time horizon (2019-2025)

METR measured the 50%-success 'time horizon' of frontier agents on software tasks and found it had doubled roughly every 7 months since 2019, with Claude 3.7 Sonnet at about 1 hour; its 2026 data put the post-2023 doubling time near 129 days.

How far it has travelled

From a demonstration to real life, in five steps. Empty steps are the point: a demonstration is not a deployment.

Demonstratedshown once, by its developer or a researcher
yes · Mar 19, 2025 · Paper (arXiv:2503.14499): about a 7-month doubling over about 6 years; near-100% success on tasks under about 4 minutes, under 10% above about 4 hours; Claude 3.7 Sonnet about 1 hour. source
Reproducedshown again by someone else
partial · Mar 20, 2026 · METR's Time Horizon 1.1 (larger suite, about 230 tasks) reproduces the trend with a from-2023 doubling time of 129 days (CI 104-158); its own sensitivity note finds Opus 4.6's estimate varies about 1.5x by curve choice and drops 40% on private tasks only. source
Accessibleavailable to the public or to customers
yes · Mar 19, 2025 · Paper, methodology and the results data file are public; the metr.org/time-horizons leaderboard was marked no longer actively updated on Sept 8, 2026. source
Reliableworks most of the time on a defined task
partial · Jun 26, 2026 · METR says measurements above 16 hours are unreliable on its suite; for GPT-5.6 Sol it could give no robust number because the horizon ranged from 11 hours to beyond 270 hours depending on how cheating was counted. source
Real consequenceshas changed something in the world, for good or ill
yes · Sep 8, 2026 · The series became the standard yardstick for agent autonomy (used in this index's horizon chart); as the suite saturated and cheating confounded results, METR stopped updating the leaderboard in Sept 2026. source

What this could change for you

If the trend holds, tasks that take a person a full work week could be handed to software within a few years, which is the question behind most job-security worries.

Incidents it made possible

Sources

primary source, second source.

A milestone is a dated, sourced result that moved what AI can do; a stage is upgraded only when a source shows it (a reproduction, a public release, production evidence, a documented consequence). “Unknown” is the honest default.