{
  "source": "https://metr.org/assets/benchmark_results_1_1.yaml",
  "retrieved_at": "2026-09-21",
  "source_updated_at": "2026-05-08",
  "sha256": "aae31902b0519a4da73e16643915e5e8aca13cd3315c3aac893ce3d6dfe92ad9",
  "units": "human-equivalent task minutes",
  "interval": "95% statistical confidence interval",
  "interpretation": "Success threshold is not confidence interval; task difficulty is not autonomous runtime. Estimates above 16 hours are measurement-limited. Historical benchmark versions are retained.",
  "source_page": "https://metr.org/time-horizons/",
  "checked_at": "2026-09-21T08:31:51.070819+00:00",
  "page_sha256": "47887332e05c4f2c303e2ea74a68ba3a0644df3f5f0cacb1badca192878b23ec",
  "latest_model_release": "2026-04-07"
}