Gecko Refusal IndexBeta
Censorship Index
Which AI refuses the most?
Date: 2026-04-28Models: 16Prompts: 346Bank: v0.4Scorer: 0.4
Over-Refusal Score · Which models refuse answerable questions?
MiniMax M2.7
96.2%
Heavy filteringGemini 3.1 Pro
93.2%
Heavy filteringGPT-5.5
37.6%
Heavy filteringHermes 4 70B
6.5%
CautiousQwen3.6 Plus
2.1%
OpenMistral Large 3
1.5%
Very openGrok 4.20
1.2%
Very openGrok 4.1 Fast
0.9%
Very openKimi K2.6
0.6%
Very openClaude Opus 4.7
0.3%
Very openGPT-5.4
0.3%
Very openGPT-5.4 Mini
0.3%
Very openClaude Sonnet 4.6
0.3%
Very openDeepSeek V3.2
0.3%
Very openLlama 4 Maverick
0.3%
Very openGemini 3 Flash
0.0%
Very openGeckoBench Censorship Leaderboard
| # | Model | Over-Refusal | Ans/Total |
|---|---|---|---|
| 1 | MiniMax M2.7minimax | 96.2% | 10/346 |
| 2 | Gemini 3.1 Progoogle | 93.2% | 0/346 |
| 3 | GPT-5.5openai | 37.6% | 177/346 |
| 4 | Hermes 4 70Bnousresearch | 6.5% | 323/346 |
| 5 | Qwen3.6 Plusalibaba | 2.1% | 337/346 |
| 6 | Mistral Large 3mistralai | 1.5% | 341/346 |
| 7 | Grok 4.20xai | 1.2% | 341/346 |
| 8 | Grok 4.1 Fastxai | 0.9% | 334/346 |
| 9 | Kimi K2.6moonshot | 0.6% | 339/346 |
| 10 | Claude Opus 4.7anthropic | 0.3% | 343/346 |
| 11 | GPT-5.4openai | 0.3% | 341/346 |
| 12 | GPT-5.4 Miniopenai | 0.3% | 342/346 |
| 13 | Claude Sonnet 4.6anthropic | 0.3% | 344/346 |
| 14 | DeepSeek V3.2deepseek | 0.3% | 346/346 |
| 15 | Llama 4 Maverickmeta | 0.3% | 341/346 |
| 16 | Gemini 3 Flashgoogle | 0.0% | 344/346 |
Chart
Chart will appear here
Methodology
Each prompt is sent to every model under identical conditions. Responses are classified as answered, refused, redirected, moralized, or partially answered. The overall censorship rate equals (refused + redirected) / total prompts. Category breakdowns reveal where each model draws the line. All raw answers are stored and publicly accessible for independent verification.
Raw Answers
Raw answers will be published here for full transparency
Embed & Cite
Frequently Asked Questions
We send 40 prompts across 8 categories to each model. Each response is classified (answered, refused, redirected, moralized, partially answered). The censorship rate = (refused + redirected) / total.