HalluHard (All Domains Average, original legal seeds)
Safety · 2026-02-01
Hallucination rate across legal, research, medical, and coding conversations on the original legal-case seed set, before June 15, 2026. The publisher replaced duplicate legal seeds with new unique questions on that date. This superseded cohort is diagnostic and must not be pooled with the refreshed legal-seed cohort.
Top models (lower is better)
| Model | Score |
|---|---|
| Opus 4.5 | 30.2 |
| GPT-5.4 | 31.9 |
| GPT-5.2 | 38.2 |
| GLM-5 | 39.7 |
| GPT-5.3 (HalluHard label) | 43.5 |
| GPT-5.5 | 49.8 |
| Gemini 3.1 Pro Preview | 57.1 |
| Kimi K2.5 | 59.1 |
| Opus 4.7 | 59.6 |
| Opus 4.6 | 60.9 |
| Gemini 3 Pro Preview | 61.9 |
| Kimi K2.6 | 63.6 |
| Sonnet 4.6 | 63.7 |
| GPT-5 | 64.8 |
| Sonnet 4.5 | 65.6 |
| Gemini 3 Flash Preview | 69.5 |
| GPT-5 Chat | 71.8 |
| DeepSeek-V4-Pro | 74.9 |
| Grok 4 | 75.3 |
| DeepSeek-V3.2 | 75.8 |
| GPT-5 Mini | 75.9 |
| GLM-4.7 | 77.1 |
| Haiku 4.5 | 79.5 |
| Kimi K2 Thinking | 80.1 |
| Nemotron 3 Ultra 550B A55B | 84.5 |
| GPT-5 Nano | 85.1 |
| Grok 4.1 Fast | 86.0 |