AI task horizons double every 7 months, METR finds

This is the Moore's Law moment for AI agency — a clean exponential curve that makes vague AI hype claims concrete and measurable.

• –METR tested frontier models across ~230 tasks, finding R²=0.83 correlation between task length and agent success — unusually tight for AI benchmarks
• –Current frontier models (e.g. Claude 3.7 Sonnet) succeed on tasks taking humans a few minutes but fail 90%+ of the time on 4-hour tasks — the "why isn't AI replacing me yet" gap explained
• –7-month doubling since 2019, now accelerating to 4 months, suggests the 2026-2027 window is when multi-day autonomous task completion becomes routine
• –The self-referential implication is stark: if agents can automate AI research, the doubling rate itself could accelerate — METR explicitly flags this flywheel risk
• –Benchmark R² of 0.83 is high but not predictive of discontinuous jumps; a single architectural breakthrough could shatter the curve in either direction