OpenAI's own reports diverge: task-length doubling slows from 2.2 to 3.8 months

scaling01 · x · 2026-11-19

scaling01 compares the 50% time horizon in OpenAI's September and November reports: the September report (Jan–Jul data) showed a 2.2-month doubling time, while the November report (Mar–Oct) shows 3.8 months. Notably, using only Mar–Jul from the September data also yields 3.8 months. He adds that GPT-6 Astra now estimates task lengths, and its estimates are much shorter than GPT-5.6 Sol's.

Original post →

More from Models

Models channel →