Factual consistency rate measured by Vectara HHEM. The inverse of hallucination rate (1 - hallucination rate). Higher is better.
Source: Vectara HHEM| Rank | Model | |
|---|---|---|
| #1 | Gemini 2.5 Pro | 93.0% |
| #2 | Meta AI Llama 4 Scout | 92.3% |
| #3 | Gemini 2.5 Flash | 92.2% |
| #4 | Gemini 3.1 Flash Lite | 91.8% |
| #5 | Meta AI Llama 4 Maverick | 91.8% |
| #6 | Anthropic Claude Haiku 4.5 | 90.2% |
| #7 | Anthropic Claude Sonnet 4 | 89.7% |
| #8 | Gemini 3.1 Pro | 89.6% |
| #9 | ChatGPT GPT-5 Nano | 89.5% |
| #10 | Anthropic Claude Sonnet 4.6 | 89.4% |
| #11 | Anthropic Claude Opus 4.5 | 89.1% |
| #12 | Anthropic Claude Opus 4.1 | 88.2% |
| #13 | Anthropic Claude Opus 4 | 88.0% |
| #14 | Anthropic Claude Sonnet 4.5 | 88.0% |
| #15 | Anthropic Claude Opus 4.6 | 87.8% |
| #16 | ChatGPT GPT-5 Mini | 87.1% |
| #17 | Gemini 3 Flash | 86.5% |
| #18 | ChatGPT GPT OSS 120B | 85.8% |
| #19 | Moonshot AI Kimi K2.5 | 85.8% |
| #20 | ChatGPT GPT-5 | 85.3% |