METR puts 5% odds on AI task horizons already exceeding 1000 hours

ben_j_todd · x · 2026-11-21

benjtodd cites METR's view that there's roughly a 5% chance model task time horizons already exceed 1000 hours. The accompanying chart from METR's report shows pooled 50% time-horizon point estimates of 40h for GPT-6 Astra and 39h for Opus 5.5, but with 95% CI upper bounds of 2645h and 1253h respectively — huge uncertainty.

Related event: METR quarterly update: AI R&D capabilities double every 1.65 months on new benchmark(12 posts)→

Original post →

More from AGI Musings

AGI Musings channel →