METR data implies AI task time horizons now double every ~2 months
scaling01 · x · 2026-11-07
METR published an analysis showing that since Oct 2025, AECI benchmark scores for Anthropic's models have grown at 1.9x [95%: 1.5–2.8] the earlier rate, while cautioning that how benchmark scores map to real-world capabilities remains unclear.
Extracting AECI scores from METR's figures, one analysis implies the task time horizon doubling time is now around 2 months, with Model 4's horizon at roughly 240 hours. The author stresses this is weak evidence — a clear 1:1 mapping from AECI to time horizon is uncertain at these magnitudes. The accompanying chart shows fitted doubling times accelerating from 260 days (2019–2023) to 92 days (Sonnet 4.5 era) to 56 days from Opus 4.5 onward.
More from AGI Musings
- Ex-OpenAI policy head: safety reports yield diminishing returns without human-level framing — deanwball · 2026-11-07
- Scott Alexander clashes with critic over whether AI labs may race without public consent — wfithian · 2026-11-07
- A 2x capabilities speedup, 3x algorithm progress: Tom Davidson says this is what 2027 IQ explosion looks like — eli_lifland · 2026-11-07
- AI labs pay up to $2M for corporate work files and emails, data trade may rival compute — pmddomingos · 2026-11-07
- Delvetown: an internet town where humans and AI agents live as neighbors hits 1,000+ residents — Miles_Brundage · 2026-11-07
- ICLR reviews flagged 91-93% AI-generated, sparking peer review integrity debate — 3scorciav · 2026-11-07