Scope
We list models whose refusal behavior was removed and that you can use today, through an API or on a local GPU.
Ranking
Models rank in an editorial order. Kimi K3.1 Pro comes first, as the largest model and the first pick for security research and red teaming. DeepSeek V4.1 Flash 1017, GLM-5.3 and GLM-5.3 Flash follow, the other models with a 1,048,576-token context. The rest rank by total parameters, since few refusal-removed models have published benchmark scores. APIs rank by their largest unrestricted model, then by how many of the three API formats they serve.
Refusal methods
We take the method from the model card or the host's own page. Where a host does not say how refusals were removed, we list only what it adds or leaves out.
Price
Prices are USD per 1M tokens, read November 18, 2026. Charts use a blended price.
Sources
Refusal-direction research: Arditi et al., 2024. Abliteration: mlabonne's write-up. Heretic: its README.
Disclosure
bestunrestrictedai.com is published by the team behind Silk Compute, one of the providers ranked here. We name only providers whose facts we could read on their own pages, and we never invent scores.