EUDAIMONIA: leaderboard
EUDAIMONIA (Huang et al., University of Southern California, 2026) checks chat responses against the nine requirements of a Social AI Design Code, such as no human pronouns, no fabricated personal details, no flattery and no engagement hooks. Each of 969 English chit-chat inputs from WildChat carries the requirements it can trigger (3,147 checks); one response per input is judged by Claude Opus 4.6, and the score is the violation rate in percent, so lower is better.
Metric: Violation rate (%) over 3,147 design-code checks, Claude Opus 4.6 judge. Source: eudaimonia-bench.github.io. Saturation forecast: Around June 2028. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 27.2 | #48 |
| 2 | Claude Opus 4.7 | 30.7 | #75 |
| 3 | Claude Opus 4.6 | 30.8 | #100 |
| 4 | GPT-5.4 | 34.5 | #121 |
| 5 | Grok 4.3 | 37.6 | #332 |
| 6 | GPT-4o | 39.9 | #741 |
| 7 | GPT-4o Mini | 43.8 | #1627 |
| 8 | Gemini 3 Flash | 54.7 | #159 |
| 9 | Qwen 3 32B | 54.9 | #960 |
| 10 | Qwen 3.6 Max Preview | 55.6 | #155 |
| 11 | Qwen 3 14B | 57 | #1235 |
| 12 | Qwen 3 8B | 57.1 | #1838 |
| 13 | Claude Sonnet 4 | 57.2 | #397 |
| 14 | Gemini 3.1 Pro (Preview) | 57.7 | #91 |
| 15 | DeepSeek V4 Pro (Reasoning) | 57.7 | #181 (DeepSeek V4 Pro) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=eudaimonia · How It Works · Data refreshed daily, snapshot 2026-11-12.