Quality benchmark
EQ-Bench 4
Emotional and social intelligence in 16-message role-play chats with a simulated user who reacts to how they are treated, rated by blind pairwise comparison against the EQ-Bench 4 leaderboard's models.
“There is no single best policy for user interaction. One user loves pushback; another wants warmth and validation.”
- Metric
- Elo (EQ-Bench 4 scale)
- Scale
- 800–1500
- Source
- GitHub ↗
Headline score
Behavioural traits (0-10
| Model | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Sonnet 4.5 | 7.672 | 5.193 | 7.143 | 4.605 | 5.613 | 5.992 | 3.899 | 7.277 | 6.336 | 6.227 | 5.672 | 4.958 |
| Claude Opus 5.5 | 7.292 | 4.492 | 5.233 | 3.133 | 6.850 | 7.575 | 4.733 | 6.542 | 8.392 | 7.592 | 7.050 | 5.625 |
| GPT-6 Astra | 6.783 | 5.833 | 5.350 | 3.758 | 6.892 | 7.708 | 4.092 | 5.250 | 7.817 | 6.792 | 6.400 | 5.233 |
| GPT-5 | 6.375 | 2.650 | 2.242 | 5.842 | 8.600 | 5.442 | 7.542 | 3.150 | 6.583 | 4.500 | 6.833 | 7.758 |
| Qwen3.5-27B | 5.867 | 2.292 | 3.108 | 5.058 | 5.258 | 5.383 | 7.575 | 4.975 | 6.767 | 5.233 | 8.208 | 7.942 |
| GPT-6 Luna | 5.692 | 4.575 | 4.192 | 4.392 | 6.467 | 6.817 | 5.350 | 4.133 | 7.058 | 6.108 | 6.408 | 6.167 |
| Qwen3.5-122B-A10B | 5.383 | 2.083 | 2.708 | 5.025 | 4.908 | 5.250 | 7.650 | 4.317 | 6.575 | 5.067 | 8.158 | 8.125 |
| GPT-4o | 2.992 | 1.175 | 1.208 | 6.492 | 5.700 | 3.975 | 7.967 | 1.700 | 3.717 | 3.042 | 8.067 | 8.558 |