OpenAI's p80/p50 time horizon ratio is ~15x, far above METR's ~4x
BasilHalperin · x · 2026-11-09
BasilHalperin highlights an anomaly: in METR data the 50% time horizon is 4x the 80% horizon, but in OpenAI's internal data the 80% horizon is 15x shorter than the 50% one (Weibull fits show H80/H50 ratios of 2.7%-8.7% across months, with shape parameter k of 0.31-0.46).
He also notes METR's own finding: when baselined against high-context humans, time horizon stops being a useful predictor of AI capability — suggesting the high/low-context ratio varies systematically across tasks, which remains puzzling.
More from Models
- Artificial Analysis launches Decision Index v2 with 45% private data and calibration-based scoring — ArtificialAnlys · 2026-11-09
- GLM-4.5 Hit #1 on Hugging Face Within 48 Hours of MIT-Licensed Release — ccerrato147 · 2026-11-09
- Has any LLM release ever been decidedly worse than its predecessor? — adonis_singh · 2026-11-09
- CEOs say Google's Argon stands out for writing, calling Google the AI dark horse — danshipper · 2026-11-09
- Epoch AI adds ECI scores, projects Gemini 3 Pro time horizon at 4.9 hours — CFGeek · 2026-11-09
- Infosec researcher maps 3 low-hanging abuse paths for open-weight models — matthew_d_green · 2026-11-09