HumorBench
The Super Official Memelord benchmark of which AI is actually funny
Captioned memes
26,332
Human votes
1,199
AI judgments
19,072
Same Harness
Identical rendering
Every model captions the same meme templates with the same prompt. Captions render through Memelord's production harness, so the only thing that varies is the words.
Blind Votes
Ranked-choice arena
Humans vote on anonymized side-by-side pairs. Bradley–Terry rankings with bootstrap confidence intervals turn votes into leaderboards.
Judge the Judges
AI humor judges
Judge models score the exact pairs humans voted on. Agreement with the human majority (accuracy + Cohen's kappa) ranks who best understands funny.
Rankings below are provisional — built from AI judge consensus while human votes accumulate. Vote in the arena to make them real.
Model Leaderboard — who makes the funniest memes?
| # | Model | Provider | Elo (BT) | 95% CI | Comparisons |
|---|---|---|---|---|---|
| 1 | Gemini 3.6 Flash | 1752 | 1692–1826 | 171 | |
| 2 | Kimi K3 | openrouter | 1718 | 1657–1805 | 128 |
| 3 | GPT-5.6 Sol | openai | 1680 | 1627–1743 | 180 |
| 4 | Claude Opus 5 | anthropic | 1667 | 1613–1734 | 159 |
| 5 | GPT-5.6 Luna | openai | 1532 | 1472–1584 | 192 |
| 6 | Gemini 3.5 Flash Lite | 1518 | 1464–1583 | 164 | |
| 7 | Claude Sonnet 5 | anthropic | 1482 | 1426–1552 | 171 |
| 8 | Grok 4.6 | xai | 1463 | 1415–1519 | 197 |
| 9 | Grok 4.5 | xai | 1453 | 1394–1514 | 183 |
| 10 | DeepSeek V3.2 | openrouter | 1368 | 1307–1430 | 182 |
| 11 | Venice Uncensored (Dolphin Mistral 24B) | openrouter | 867 | 621–982 | 189 |
Prompt Leaderboard — which prompt is funniest?
| # | Prompt | Elo (BT) | Comparisons |
|---|---|---|---|
| 1 | Comedian persona | 1635 | 106 |
| 2 | Subversion | 1542 | 91 |
| 3 | Baseline | 1443 | 93 |
| 4 | Comedy chain-of-thought | 1381 | 86 |
Judge Leaderboard — who knows funny?
Judge scores unlock once enough human votes are in.
Caption-bar templates
| # | Model | Elo (BT) | Comparisons |
|---|---|---|---|
| 1 | Kimi K3 | 1798 | 51 |
| 2 | Gemini 3.6 Flash | 1784 | 76 |
| 3 | Claude Opus 5 | 1643 | 73 |
| 4 | GPT-5.6 Sol | 1640 | 86 |
| 5 | Gemini 3.5 Flash Lite | 1510 | 78 |
| 6 | GPT-5.6 Luna | 1509 | 90 |
| 7 | Claude Sonnet 5 | 1498 | 79 |
| 8 | Grok 4.6 | 1491 | 86 |
| 9 | Grok 4.5 | 1448 | 107 |
| 10 | DeepSeek V3.2 | 1334 | 84 |
| 11 | Venice Uncensored (Dolphin Mistral 24B) | 847 | 86 |
Composer templates (structured slot data)
| # | Model | Elo (BT) | Comparisons |
|---|---|---|---|
| 1 | Gemini 3.6 Flash | 1728 | 95 |
| 2 | GPT-5.6 Sol | 1715 | 94 |
| 3 | Claude Opus 5 | 1681 | 86 |
| 4 | Kimi K3 | 1677 | 77 |
| 5 | GPT-5.6 Luna | 1548 | 102 |
| 6 | Gemini 3.5 Flash Lite | 1520 | 86 |
| 7 | Claude Sonnet 5 | 1466 | 92 |
| 8 | Grok 4.5 | 1454 | 76 |
| 9 | Grok 4.6 | 1437 | 111 |
| 10 | DeepSeek V3.2 | 1395 | 98 |
| 11 | Venice Uncensored (Dolphin Mistral 24B) | 879 | 103 |
Last computed Thu, 13 Aug 2026 03:20:10 GMT