HalluHard (All Domains Average, June 2026 legal refresh)
Safety · 2026-06-15
Hallucination rate across legal, research, medical, and coding conversations after the June 15, 2026 replacement of duplicate legal-case seed questions with unique questions. Only runs explicitly marked by the publisher as evaluated on the refreshed benchmark belong to this cohort.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 27.2 |
| Claude Fable 5 | 38.9 |
| Sonnet 5 | 59.7 |
| GLM-5.2 | 74.8 |