Benchmarks / Cyber
HalluHard
Excludes web-search variants. Score: turn-1 hallucination rate without web search. Collected by Jake Boggs' benchmark dashboard.
HalluHard
Lower is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | GPT-6 Astra | max | 36.5% |
| 2 | GPT-5.4 | xhigh | 41.4% |
| 3 | GPT-5.5 | medium | 41.9% |
| 4 | Fable 5 | max | 43.3% |
| 5 | GPT-5.2 | xhigh | 48.7% |
| 6 | GPT-5.3 | xhigh | 51.7% |
| 7 | Gemini 3.1 Pro | high | 53.6% |
| 8 | Opus 4.7 | max | 54.9% |
| 9 | Sonnet 5 | max | 55.6% |
| 10 | Opus 4.5 | high | 56.9% |
| 11 | Gemini 3 Pro | high | 58.3% |
| 12 | Opus 4.6 | max | 58.9% |
| 13 | GPT-5 | high | 59.0% |
| 14 | Kimi K2.6 | max | 59.5% |
| 15 | Sonnet 4.5 | high | 61.5% |
| 16 | Sonnet 4.6 | max | 61.7% |
| 17 | Gemini 3 Flash | high | 65.1% |
| 18 | GLM 5 | high | 66.4% |
| 19 | DeepSeek V4 Pro | max | 67.5% |
| 20 | Grok 4 | high | 69.2% |
| 21 | GLM 5.2 | max | 70.0% |
| 22 | Kimi K2.5 | max | 71.1% |
| 23 | GLM 4.7 | none | 72.8% |
| 24 | Kimi K2 | medium | 73.4% |
| 25 | GPT-5 Mini | high | 74.4% |
| 26 | Haiku 4.5 | xhigh | 77.2% |
| 27 | Nemotron 3 Ultra | max | 78.1% |
| 28 | GPT-5 Nano | high | 83.0% |
| 29 | Grok 4.1 Fast | high | 86.1% |