How long can AI work alone? The ruler just ran out.
METR measures how long a task an AI agent can finish on its own, in the time it would take a skilled person.
The numbers
- METR, March 20257 mo
for the task length AI can handle to double.
Measured over 2019 to 2025. ‘Length’ means how long the task takes a skilled person.
Source: METR, 19 Mar 2025
- METR, January 2026 update89 days
is the doubling time since 2024.
88.6 days. METR says part of the change comes from re-estimating models: older ones down, recent ones up.
Source: METR, Time Horizon 1.1, 29 Jan 2026
- 50% time horizon: task length finished half the time
From 2 hours to 16+ hours
- GPT-5, Aug 20252 hr 17 min
- Claude Mythos Preview (early), May 202616+ hrs
Mythos: early version, 95% interval of 8.5 to 55 hours.
Source: METR via The Decoder; METR time-horizons page
- Why METR can’t measure higher5 of 228
tasks in its suite take 16 hours or more.
METR: measurements above 16 hours are unreliable with its current tasks.
Source: METR time-horizons page; The Decoder
- Read it carefully
"50%" means it fails half the time
A 16-hour horizon means the agent finishes 16-hour tasks about half the time, on METR’s kind of tasks. Not that it works reliably for two days.
On real freelance projects, success rates don’t track this curve well. Treat it as a trend line, not a promise.
Source: METR; CAIS, 1 Jul 2026
Give agents longer tasks, with checkpoints
- Hand over whole tickets, not single functions. The trend says the next model handles more than you’re giving it.
- Add checkpoints. At a 50% success rate, a review every hour or two saves the day.
- Re-test every model release. A task that failed in spring may pass now.
Sources
- METR, Measuring AI ability to complete long tasks, 19 Mar 2025
- METR, Time Horizon 1.1, 29 Jan 2026
- METR, Time horizons page, 2026
- The Decoder, citing METR, 10 May 2026
What’s the longest task you’d give an agent today? Tell me on LinkedIn or Instagram.