The Measure of an Agent's Power

How long can it work autonomously?

METR time horizon exponential trend

Claude Opus 4.5 ~5 hours

GPT-5.1 Codex Max ~3 hours

Claude Sonnet 4.5 ~2 hours

Gemini 3 Pro ~2 hours

@ 50% success rate

~7 months

Doubling time

6+ years

Exponential trend

Minutes → Hours (NOW) → Days → Weeks

Source: METR Time Horizons 1.1 (Jan 2026)