← Benchmark report cards
Measured benchmarks
Humanity's Last Exam
Published by Scale AI / CAIS. Source page ↗ · Data we read (file download) ↗
Scale AI / CAIS; a public set plus a held-out set.
Facts
| Fact | Value | Where it comes from |
|---|---|---|
| Who runs the evaluation | an independent organisation | curated by aiqbench, reviewed 2026-10-09 |
| Test set exposure | part public, part held out | curated by aiqbench, reviewed 2026-10-09 |
| Data licence | CC-BY-4.0 | the board's attribution record |
| Newest published result | Oct 7, 2026 | the ingested rows |
| Last pulled by aiqbench | 21 minutes ago · every 12 h | the sync log |
| Models on this axis | 56 | the ingested rows |
| Rows with a published interval | every row | the ingested rows |
| Headroom at the top | Best 54.8%, 45.2% below the ceiling; top 10 within 18.6%. | computed from the ingested rows |
| Axes it tracks closely |
| Spearman ρ over shared models (see Clusters) |
Last frontier jumps
- Sep 13, 2026 · gpt-6-astra · 54.8%
- Sep 13, 2026 · claude-fable-5-1 · 46.5%
- Mar 5, 2026 · gpt-5.4-pro-2026-03-05 · 44.3%
Where it appears
Weighted in the built-in templates: All sources ×1, Reasoning ×1.
View this axis alone in the console →Attribution
Source: Epoch AI, 'AI Benchmarking Hub'. Published online at epoch.ai. Retrieved from https://epoch.ai/benchmarks. Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Modified by aiqbench: normalised into aiqbench's common board format; we ingest the 'hle_external.csv' table and rescale its 0-1 scores to percent. Underlying results published by Scale AI / CAIS — https://agi.safe.ai/, transcribed by Epoch AI.