Atlas · C14 · Core cognition
Persistence on long tasks
Carrying a single piece of work through to the end over hours, days or weeks.
AI Reliable score
4.6 / 10
Average adult level · peak 6.7 · trend ↑↑ · grade A
How it is calculated
METR 80% time horizon ≈ 3.1 h → H-scale (1 h = 4, 1 workday = 6, 1 week = 8)
minus 0.5 reliability margin = 4.6 Reliable
Peak from the 50% horizon (≈ 17.4 h) = 6.7
Why this score
Computed on the Persistence sheet from METR time horizons: Reliable = 80% horizon (~3 h) minus a 0.5 discount because METR tasks are self-contained software work; Peak = 50% horizon (~17 h). Doubling roughly every 4 months.
On the human scale
2Beginner. ~5-minute tasks
4Average adult. ~1-hour tasks ← AI today
6Competent pro. ~1-workday tasks
8Expert. ~1-work-week tasks
10Best humans. Multi-month projects
Future warning
At its best, AI already reaches 6.7 on this capability. Peak results are never counted in the index; they show where the water may rise next.
Evidence
METR time-horizon measurements (via Wikipedia and AGI Tracker summaries)A
E01 · 2026-02 to 2026-09
- Measures
- Length of software tasks (in human-expert time) an AI agent completes at 50% / 80% reliability
- AI result
- Claude Mythos Preview: ~17.4 h at 50%, ~3 h 06 min at 80%. Claude Opus 4.6: ~12 h / ~70 min. Doubling ~4.3 months (since 2023).
- Humans
- Human expert time (the metric itself)
- On our scale
- Converted to the H-scale on the Persistence sheet (1 h = 4, 1 workday = 6, 1 week = 8). Caveats: self-contained software tasks only; computer-use horizons 40-100x shorter; >16 h unreliable; at least 16% of long 'successes' were cheating.
en.wikipedia.org ↗
agitracker.io ↗
History
First reading: Atlas v0.1, 5 Nov 2026. Every later change will be listed here with its reason.
Disagree? Send a correction
