A price tag and a report card for uncensored AI.

The problem

Open-weight models keep getting better, and removing their refusal behavior is now a routine, well-documented technique. A small market of hosted services has grown up around that: chat apps and APIs that promise models which simply answer.

Comparing them is hard. Prices are quoted in different units, model names hide what they're built on, "no logs" means different things to different companies, and claims about being uncensored are rarely tested.

What we do

• Catalog every hosted provider we can find, with models, prices, and policies linked to their sources.
• Measure capability and refusal behavior ourselves, against each provider's live endpoint.
• Normalize price to USD per million tokens so you can compare like with like.
• Publish the method, the weights, and the dates, so every number can be checked.

How we research

• Primary sources only. Prices, models, and policies come from each provider's own pricing pages, model APIs, privacy policies, and terms. Every provider page lists the sources and the date we checked them.
• Documented claims, not marketing. Privacy scores only count what a policy actually says. If a provider doesn't write it down, it doesn't get the points.
• Fact-checked before publishing. Every review and article is checked line by line against our data and the sources it cites. When a provider's page changes, we update the data and every post that relies on it.
• No testing claims we haven't made. Until our benchmark runs are complete, reviews describe what providers publish, and say so.

Independence

We're not affiliated with any provider. Placement in the index can't be bought. If we ever use affiliate links, they'll be labeled and won't affect rankings.

Corrections and submissions

Spotted a wrong price, a new provider, or a policy change? Email[email protected] with a link to the source.