The longest task an AI can finish on its own

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]


1,045 minutes

Held by claude-mythos-preview-early. Set 2026-04-07, and has stood for 175 days. Written down here 2026-09-07.

Give a model a job, and time how long the same job takes a competent human. Models finish short jobs and fail long ones, so the length at which a model succeeds half the time is a single number describing how much work it can be left alone with. In 2019 that number was six seconds. It has been doubling roughly every four months ever since, and it has now passed seventeen hours, which is longer than a working day. The measure comes from METR, which evaluates models for exactly this purpose. It says nothing about whether the work was any good, only that it was finished. That is a real limitation and it is still the closest thing anyone publishes to a speedometer for the field.

That description was written when this frontier was opened, on 2026-09-07, and is never rewritten. Where it and the number above it disagree, the number is the reading and the description is history.


What counts
The length of job, measured in how long it takes a competent human, that a model completes successfully half the time. METR's figure, at the fifty per cent mark, for whichever model it has marked as the state of the art - including preview releases, which METR marks alongside the rest.
What moves it
METR evaluates a new model and it beats the standing figure.

Where to check it: https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/

The program reads it from https://metr.org/assets/benchmark_results_1_1.yaml, which is the same figure in a form a script can parse.


Last checked: 2026-09-28


Previous: Assembly mistakes spotted in a photograph   Next: Open problems in mathematics solved

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]