METR: AI task-completion capability doubles every 3.6 months, 200h tasks by mid-2027
141_1337 · reddit · 2026-11-21
A new METR study tracks frontier models' task-length capability: the time horizon of tasks AI can complete doubles roughly every 3.6 months (95% CI 2.08–4.94 months).
- The curve spans GPT-4o (5 minutes) through o1, GPT-5, Opus 4.5, Opus 4.6 and Fable 5, up to GPT-6 Astra approaching 200-hour tasks.
- Both TH1 and TH2 time horizons show steady exponential growth; longest tasks measured are 30h (TH1) and 200h (TH2).
- On easier-to-verify tasks progress may be faster, with AI potentially reaching 200-hour human-level tasks by July 2027.
- Bootstrap distribution of doubling times clusters tightly around 3.6 months, indicating the trend is robust across model families.
More from AGI Musings
- Claim: a proof of the Riemann Hypothesis found as Metaculus forecast leaps from 2044 to 2035 — haider1 · 2026-11-21
- Thiel doubts AI lab monopolies, xAI rebuild began when Musk showed up, and more signals — himanshustwts · 2026-11-21
- Terence Tao: AI is winning at math but we've learned almost nothing mathematically — GaryMarcus · 2026-11-21
- Smaller experiments win: the Kaggle lesson for doing more AI research with less compute — strickvl · 2026-11-21
- Ramez Naam extrapolation: AI a decade away from completing 98% of hour-long human tasks — GaryMarcus · 2026-11-21
- When should an AI override self-destructive user requests? A paternalism debate — LookOverall · 2026-11-21