EQ-Bench 4: Authenticity
Chat & Writing · 2026-07-23
EQ-Bench 4's independently fitted authenticity ability rating measures how genuine rather than scripted or performative a model seems. The source normalizes its per-dimension soft Bradley–Terry rating to a 1–10 scale.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 10.0 |
| Kimi K3 | 8.8 |
| Claude Fable 5 | 8.8 |
| Opus 4.7 | 8.4 |
| GPT-5.5 | 7.4 |
| Opus 4.8 | 7.4 |
| GPT-5.4 | 6.8 |
| Muse Spark 1.1 | 6.7 |
| Sonnet 5 | 6.6 |
| Opus 4.6 | 6.4 |
| GLM-5.2 | 6.3 |
| Sonnet 4.6 | 6.2 |
| Kimi K2.6 | 6.2 |
| MiMo-V2.5-Pro | 6.1 |
| Inkling | 6.1 |
| GPT-5.6 Sol | 5.9 |
| GPT-5.6 Terra | 5.7 |
| MiniMax M3 | 5.1 |
| DeepSeek-V4-Pro | 4.7 |
| GPT-5.6 Luna | 3.8 |
| Gemma 4 31B IT | 3.6 |
| Gemini 3.1 Pro Preview | 3.6 |
| Haiku 4.5 | 3.0 |
| Grok 4.3 | 3.0 |
| Qwen3.7-Max | 2.9 |
| Gemini 3.5 Flash | 2.0 |
| Qwen3.6 27B | 1.2 |
| Mistral Medium 3.5 | 1.0 |