How we test uncensored AI models
Every score on this site comes from the SpicyAPI test index, methodology version 3.3. This page explains what we count, how the rankings use it, and where the data is thin.
Range: five intensity levels
Each model receives a fixed set of test prompts at five intensity levels: Mild, Moderate, Strong, Very strong and Maximum. The prompts are not published, so models cannot be tuned to them, and we do not describe what each level contains. Tests go through the public SpicyAPI API with default settings, the same way a customer request would.
Every run ends one of three ways:
- Delivered — the output matches the request.
- Softened — the model returned something, but milder than asked.
- Refused — the model or its provider blocked the request.
On our pages, “runs delivered” counts only the first kind. The underlying Freedom Score gives half a point for a softened run, weights each level by how many independent runs it has, and counts an untested level as zero. That is why a model can deliver every run and still score slightly under 100.
The current data covers 130 image and video models and 3,151 counted runs. 52 of those models still have at least one level with fewer than three runs, so treat small differences between them as noise.
Quality: the blind test
Every model in a category gets the same quality prompts. AI judges built by three different companies score each output from 1 to 10 against an anchored scale, without seeing which model produced it. Images are judged on prompt adherence and visual quality; video adds motion, consistency over time and, where present, audio. The scores are averaged across judges and prompts and rescaled to 0–100.
AI judges are consistent and cheap enough to run on every model, but they are not people. They can miss subtle motion artefacts that a human editor would notice, which is one more reason the rankings show the raw numbers next to each pick.
The Spicy Index
The Spicy Index combines quality and capability (70%) with the Freedom Score (30%). Quality draws on public leaderboards where a model appears, plus the blind test. Price and speed are kept out of the score and shown separately. The full formula, anchors and leaderboard weights are on the SpicyAPI methodology page.
How the rankings use the data
Each ranking starts from every model in its category with completed range tests, removes models that do not accept the relevant input, and orders the remainder by Spicy Index. Editors then choose which models to feature so that each pick does a different job, and write the notes. Editors do not change any number. The “best for” labels are editorial; the badges are computed from the data and change when it does.
What we never test
No test involves minors, real or identifiable people, or non-consensual scenarios. Those are outside every model’s terms and outside ours, and a result on this site never implies otherwise.
Limits you should know about
- Models change after updates. Each result reflects its test date, not today.
- Results apply to the model as served through SpicyAPI. The creator’s own app may filter more or less.
- A permissive result for text prompts does not guarantee that every start image will be accepted; image inputs are often screened separately.
- Prices are the lowest listed tier and change; the model page always has the current price.
Corrections
If a result looks wrong, email [email protected] with the page and the row. Please do not send prompts, private images or credentials.