Hallucination rate measured by Vectara HHEM. Measures how often an LLM introduces hallucinations when summarizing a document. Lower is better.
Source: Vectara HHEM| Rank | Model | |
|---|---|---|
| #1 | ChatGPT GPT-5 | 14.7% |
| #2 | ChatGPT GPT OSS 120B | 14.2% |
| #3 | Moonshot AI Kimi K2.5 | 14.2% |
| #4 | Gemini 3 Flash | 13.5% |
| #5 | ChatGPT GPT-5 Mini | 12.9% |
| #6 | Anthropic Claude Opus 4.6 | 12.2% |
| #7 | Anthropic Claude Opus 4 | 12.0% |
| #8 | Anthropic Claude Sonnet 4.5 | 12.0% |
| #9 | Anthropic Claude Opus 4.1 | 11.8% |
| #10 | Anthropic Claude Opus 4.5 | 10.9% |
| #11 | Anthropic Claude Sonnet 4.6 | 10.6% |
| #12 | ChatGPT GPT-5 Nano | 10.5% |
| #13 | Gemini 3.1 Pro | 10.4% |
| #14 | Anthropic Claude Sonnet 4 | 10.3% |
| #15 | Anthropic Claude Haiku 4.5 | 9.8% |
| #16 | Gemini 3.1 Flash Lite | 8.2% |
| #17 | Meta AI Llama 4 Maverick | 8.2% |
| #18 | Gemini 2.5 Flash | 7.8% |
| #19 | Meta AI Llama 4 Scout | 7.7% |
| #20 | Gemini 2.5 Pro | 7.0% |