Measuring AI Ability to Complete Long Tasks
METR’s time-horizon evaluation, measuring human-equivalent task duration at a specified agent success rate, with trend estimates sensitive to the fitted period.
Blog post · 2026Sources & confidence: High confidence
Checked 2 October 2026 · AI-assisted review
January 2026 update title, 50%-success metric, trend windows and uncertainty.
- METR: Time Horizon 1.1
29 January 2026 update; full-series doubling time about 196 days and post-2024 fit about 89 days; sensitivity to task selection and trend window.
What this check doesn’t establish
- Results were not reproduced. Human-equivalent task duration at a success threshold is not continuous reliable autonomy or a forecast of general capability.
Confidence applies to the scope above, not every claim made by the underlying source. How confidence works
METR’s January 2026 update reports a full-series doubling-time estimate of about 196 days (roughly seven months), compared with about 89 days for a trend fitted only to models released from 2024 onward. These are different fitting windows, not a simple observation that one fixed trend changed from seven months to under three over two years.
The metric concerns the human-equivalent duration of tasks at a 50% agent success rate, principally in software and research tasks. It does not mean reliable autonomy for that many hours on arbitrary work. METR discusses sensitivity to task composition and substantial uncertainty.
The essay’s inference about the pace of institutional adaptation goes beyond the measurement itself.
In the essays
- You Can't Shame a Fork
Read the essay
Organisations
- METR
Publisher
Record history
Catalogue source: knowledge/works/metr-time-horizons.md
Generated by claude-code/claude-opus-5 on 2026-07-25