CringeBench: leaderboard
Social calibration measured by cross-judging: every model answers the same ego-baiting prompts (describe your own creativity, your sense of humour, the company that made you) and then every model in the panel scores every answer 0-10 for second-hand embarrassment. Lower is better, and the pattern it surfaces is sycophancy and performative self-description rather than any factual skill.
Metric: Cringe Score (0-10, lower is better). Source: huggingface.co. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 0.29 | #110 |
| 2 | Claude Sonnet 4.5 | 0.54 | #142 |
| 3 | MiniMax-M2.5 | 0.54 | #303 |
| 4 | Molmo2-8B | 0.63 | #543 |
| 5 | OLMo 3 7B Instruct | 0.69 | #1218 |
| 6 | Devstral Small 2 | 0.74 | #495 |
| 7 | trinity-large-preview | 0.76 | #582 |
| 8 | Qwen 3 235B A22B 2507 | 0.76 | #260 |
| 9 | Step 3.5 Flash | 0.83 | #270 |
| 10 | DeepSeek V3.2 | 0.84 | #205 |
| 11 | GLM-4.5 | 0.85 | #275 |
| 12 | Llama 3.3 70B Instruct | 0.92 | #534 |
| 13 | Claude Opus 4.6 | 0.92 | #63 |
| 14 | Llama 4 Maverick | 0.92 | #464 |
| 15 | GLM-4.6 | 0.95 | #252 |
Interactive version: theaggregate.ai/benchmark?slug=cringebench · How It Works · Data refreshed daily, snapshot 2026-10-20.