EQ-Bench 4

Citation frequency 2/3
What does it measure?
Tests how well the AI reads and responds to the other person's feelings in emotionally difficult conversations, judging six aspects such as rapport, authenticity and attunement.
Who evaluates it, and how?
The model holds a 16-turn role-play chat with each of 120 simulated personas. Three AI judges compare two models' conversations side by side and pick the better one, and the results become an Elo score.
Reading the score
Scores are Elo ratings from head-to-head verdicts by AI judges. Higher is better.
Top score1365
Models tested30
Scoring basisHigher is better
Last updated2026.10.24
Model release date

Can it read people's feelings and respond well in hard conversations?

EQ-Bench 4

  1. 🥇AnthropicClaude Opus 5148114861365
  2. 🥈AnthropicClaude Opus 5.5151615011340
  3. 🥉MetaMuse Wonder149715221325
  4. #4Moonshot AIKimi K3147515021319
  5. #5AnthropicClaude Fable 5150915171313
  6. #6OpenAIGPT-5.5146614841295
  7. #7AnthropicClaude Opus 4.7148315141293
  8. #8OpenAIGPT-6.1 Sol147014871272
  9. #9AnthropicClaude Opus 4.8146814931268
  10. #10OpenAIGPT-5.4146314931258
  11. #11MetaMuse Spark 1.1145814971244
  12. #12GoogleGemini 4 Argon152115421236
  13. #13OpenAIGPT-5.6 Sol148314911236
  14. #14AnthropicClaude Sonnet 5145114731221
  15. #15OpenAIGPT-5.6 Terra144714741220
  16. #16AnthropicClaude Opus 4.6148915091210
  17. #17Z.aiGLM 5.2145914771208
  18. #18XiaomiMiMo V2.5 Pro144914801198
  19. #19AnthropicClaude Sonnet 4.6145514801197
  20. #20Moonshot AIKimi K2.6144114621191
  21. #21DeepSeekDeepSeek V4 Pro145114751156
  22. #22OpenAIGPT-5.6 Luna143314591146
  23. #23MiniMaxMiniMax M3141014471140
  24. #24GoogleGemini 3.1 Pro148214951132
  25. #25GoogleGemma 4 31B143414651110
  26. #26AlibabaQwen3.7 Max145814801100
  27. #27GoogleGemini 3.5 Flash147214791077
  28. #28SpaceXSIGrok 4.3142414471065
  29. #29AnthropicClaude Haiku 4.5139914251053
  30. #30Mistral AIMistral Medium 3.514041432983
Half of models ≤ 1215