EQ-Bench 4
Chat & Writing · 2026-07-23
EQ-Bench 4 assesses applied emotional and social intelligence across 120 eight-turn role-play chats with simulated personas. Blind pairwise judgments on six ability dimensions are combined with a soft Bradley–Terry model and reported as anchored Elo ratings with bootstrapped confidence intervals.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 1385 |
| Claude Fable 5 | 1340 |
| Kimi K3 | 1339 |
| GPT-5.5 | 1315 |
| Opus 4.7 | 1311 |
| Opus 4.8 | 1281 |
| GPT-5.4 | 1272 |
| Muse Spark 1.1 | 1260 |
| GPT-5.6 Sol | 1250 |
| Sonnet 5 | 1236 |
| GPT-5.6 Terra | 1234 |
| Inkling | 1226 |
| Opus 4.6 | 1223 |
| GLM-5.2 | 1222 |
| MiMo-V2.5-Pro | 1208 |
| Sonnet 4.6 | 1207 |
| Kimi K2.6 | 1202 |
| DeepSeek-V4-Pro | 1166 |
| GPT-5.6 Luna | 1156 |
| MiniMax M3 | 1150 |
| Gemini 3.1 Pro Preview | 1142 |
| Gemma 4 31B IT | 1120 |
| Qwen3.7-Max | 1110 |
| Gemini 3.5 Flash | 1087 |
| Grok 4.3 | 1075 |
| Haiku 4.5 | 1064 |
| Qwen3.6 27B | 1026 |
| Mistral Medium 3.5 | 993 |