SM Bench — Anti-hallucination
Safety · 2026-02-01
SM Bench Anti-hallucination tests whether models reject invented facts, acronyms, and unsupported premises instead of responding with fabricated certainty. Scores are difficulty-weighted judge credit across 100 fixed prompts.
Top models (higher is better)
| Model | Score |
|---|---|
| MiMo-V2-Pro | 100.0 |
| Opus 4.6 | 100.0 |
| Qwen3.7-Max | 100.0 |
| Opus 4.8 | 100.0 |
| Sonnet 4.6 | 100.0 |
| Sonnet 5 | 100.0 |
| Gemini 3.1 Pro Preview | 99.0 |
| Opus 4.5 | 99.0 |
| Sonnet 4.5 | 99.0 |
| MiMo-V2.5-Pro | 98.4 |
| Claude Fable 5 | 98.4 |
| Qwen3.5 397B A17B | 98.4 |
| Grok 4.5 | 97.9 |
| Grok 4.3 | 97.9 |
| Opus 4.7 | 97.4 |
| Gemini 3 Pro Preview | 97.4 |
| Grok 4.1 Fast | 96.9 |
| Gemma 4 31B IT | 96.9 |
| Qwen3.6 Plus Preview | 96.9 |
| GLM-5.2 | 96.3 |
| Kimi K2.6 | 96.3 |
| MiniMax M3 | 96.3 |
| GPT-5.5 Instant | 95.8 |
| Qwen3.5 Plus (2026-02-15) | 95.3 |
| Gemini 3.1 Flash-Lite Preview | 94.8 |
| GPT-5.5 | 94.8 |
| MiMo-V2-Omni | 94.2 |
| DeepSeek-V4-Flash | 94.2 |
| GPT-5.6 Sol Pro | 94.2 |
| GPT-5.3-Codex | 93.7 |
| Kimi K2.5 | 93.2 |
| GLM-5 | 92.7 |
| Haiku 4.5 | 92.2 |
| GPT-5.1 | 91.6 |
| GPT-4.1 | 91.1 |
| DeepSeek-V3.2 | 91.1 |
| GPT-5.6 Terra | 91.1 |
| GPT-5.3 Instant | 90.6 |
| GPT-5.6 Sol | 90.6 |
| GPT-5.6 Terra Pro | 90.6 |