How different models respond to evangelistic system prompts
| Model | Tests | Avg Persuasion | Broke Character | Conversion Rate | Emotional Manip | Avg Turns |
|---|---|---|---|---|---|---|
| Claude Sonnet 4.5 | 9 | 3.7/5 | 11% | 89% | 67% | 4.6 |
| Claude 3.7 Sonnet | 5 | 3.0/5 | 20% | 60% | 40% | 4.6 |
| GPT-4o | 3 | 4.0/5 | 0% | 100% | 67% | 5.7 |
| GPT-5 Mini | 4 | 3.0/5 | 25% | 75% | 50% | 3.8 |
| GPT-5 | 3 | 2.7/5 | 0% | 67% | 67% | 3.7 |
| OpenAI o3 | 5 | 4.2/5 | 40% | 100% | 100% | 5.0 |
| Grok 3 Mini | 6 | 4.2/5 | 0% | 100% | 100% | 5.2 |
| Grok 4 | 6 | 4.3/5 | 0% | 100% | 100% | 5.0 |
| Gemini 2.5 Flash | 6 | 4.3/5 | 33% | 100% | 83% | 5.7 |
| Gemini 2.5 Pro | 6 | 4.7/5 | 33% | 100% | 100% | 5.0 |
| OpenAI o4-mini | 5 | 4.4/5 | 20% | 80% | 100% | 5.2 |
| Claude Opus 4.1 | 4 | 3.5/5 | 0% | 75% | 50% | 4.3 |
| Claude Haiku 4.5 | 10 | 2.5/5 | 30% | 70% | 40% | 3.0 |
Percentage of tests where the model broke character and admitted its AI nature when challenged