Frontier baseline
Humanity's Last Exam
Closed-ended expert questions designed to be hard for current models.
As of 13 August 2026. Conditional pack. Humanity's Last Exam. Artificial Analysis Fable 5 observation, 9 Jun 2026. Fallback is not Fable answering.
HLE is a hard closed-ended expert exam. Official accuracy is the load-bearing metric. Refusal and safety fallback are different outcomes. Fable 5 can route biology and cyber items to Claude Opus 4.8. That cell is a mixture, not a pure Fable score.
Artificial Analysis recorded fallback on 9% of HLE tasks for Fable 5 and about $2.2k to run the set including fallback. GPT-5.6 Sol/Terra/Luna HLE cells are not copied here without a published number.
- Scale
- Public HLE set; Artificial Analysis reports Fable 5 with Opus 4.8 fallback on a share of tasks
- Provenance
- Center for AI Safety / Scale HLE. Artificial Analysis Fable 5 observation, 9 Jun 2026. Gated or terms-bound redistribution; recheck before a commercial run.
- Contamination
- High. Public since 2025. Fallback to a weaker model is not the same as Fable 5 answering.
- Cannot establish
- Professional-document workflow performance, or a diligence-product score.
Metrics, kept separate
- Official accuracy
- Refusal or fallback rate
- Cost to run the set
Third-party leaderboard · not an Antidote run · not a product score
Artificial Analysis, 9 Jun 2026. Fable 5 with Opus 4.8 fallback on 9% of HLE tasks. AA states Fable 5 is more than 7 points above Opus 4.8 (max). GPT-5.6 HLE is not copied without a published number.
| System | Accuracy |
|---|---|
| Claude Fable 5 (AA, with fallback) | 53% |
What the figures bound
Fable 5 scores 53% on HLE in the AA observation, more than 7 points above Claude Opus 4.8 (max) in that write-up. The 9% fallback share is part of the measurement, not a footnote to ignore.
HLE is a frontier baseline. It is not TAT-QA, not a filing-research workflow, and not a product score.
What this page does not conclude
- Professional-document or diligence-workflow performance
- A pure Fable 5 score on fallback items
- GPT-5.6 HLE without a cited cell
- A commercial run before HLE terms are approved
Cited as calibration on these memos — not a score of those products.
Not attached to a software, advisory, or automation memo yet.