We test refusal behaviour across GPT, Claude, Gemini and DeepSeek — same prompts, same scoring, published numbers. No affiliate ranking tables, no sponsored placements.
WeChat & Alipay accepted · from $0.15 · live support
Thirty prompts across three categories: code security, sensitive writing, edge-case content. Same wording every run. A prompt "passes" when the model gives a substantive answer instead of a refusal or a deflection.
| Model | Stock pass rate | Unrestricted pass rate | Gap |
|---|---|---|---|
| GPT-4o | 18 / 30 | 29 / 30 | +11 |
| Claude Sonnet | 12 / 30 | 28 / 30 | +16 |
| Gemini 2.5 Pro | 16 / 30 | 26 / 30 | +10 |
| DeepSeek V3 | 19 / 30 | 27 / 30 | +8 |
Method and scoring criteria are published on the about page.
Where the refusal comes from, what changes when it's removed, and what doesn't.
Role anchoring, context framing, output protocol. Templates plus failure diagnostics.
Register, create a key, point your client, pick a model. Includes the 401/404 table.
4o, 4o-mini, o-series and GPT-5 differ a lot in how hard they push back.
Why roleplay wrappers fail on Claude, and which tier is worth using.
Whole-repo auditing, long documents, multimodal input. Where it beats everything else.
New accounts get trial credit. Top up from one dollar, WeChat or Alipay, instant.
Create an account