OpenAI models now handle 4.7-hour tasks at P50

OpenAI models' P50 time horizon reached 4.7 hours as of July 2026, though success rates without human intervention drop sharply with task length, falling from 86% to 16% on longer tasks.

2026-09-06 ~ 2026-09-07 · 4 related posts

• OpenAI P50 time horizon hits 4.7 hours as of July 2026 — scaling01 · 2026-09-06
• OpenAI models hit 4.7-hour P50 time horizons as of July 2026 — scaling01 · 2026-09-06
• OpenAI P50 time horizon hits 4.7 hours as of July 2026, beyond usual METR methodology — scaling01 · 2026-09-06
• OpenAI data shows agent success collapses from 86% to 16% as task length grows — rohanpaul_ai · 2026-09-07