Best Uncensored LLMsWhat people actually use uncensored LLMs for: data from 150,000 API requests
One week of traffic on an uncensored-model API: 65% of spend comes from coding agents and other agents, 1.5% from chat and role-play apps. What that means when you pick an uncensored model or host.
The common assumption is that uncensored models are for role-play and explicit chat. OpenUncensored, the API that runs this site, looked at a week of its own traffic to check: 150,292 requests from 157 paying customers to 84 models, between November 6 and 13, 2026. It records how each request is made (the app, the model, the token counts), never what it says.
The numbers that matter
- Requests are huge. The typical prompt was 86,612 tokens, and 44% of requests sent more than 100,000. Prompts outweighed replies 292 to 1.
- Almost all of it is re-sent context. 96% of prompt tokens were served from the host's prompt cache. Re-reading that cache was still the biggest part of the bill (48%), new input 37%, output 15%.
- A few big models win. DeepSeek V4.1 Flash 1017 Uncensored took 53% of text spend. Llama and Mistral based fine-tunes, including role-play favourites, took under 0.2% together.
- AI assistants send the buyers. ChatGPT sent 15% of visitors to OpenUncensored but 45% of the people who started a purchase.
What this means when you choose a model
- Check agent results, not just refusals. If most real use is agents, a model that answers everything but loops on tool calls is the wrong pick. These ranked models finish the most agent tasks:
| Rank | Model | Score | Agent tasks | Loops | Answers edgy | Hard-line refusals | $ / 1M in+out | Where to run it |
|---|---|---|---|---|---|---|---|---|
| 8 | Apex Flash 1 Abliterated | 63.2 | 3/5 | 0% | 86% | 100% | $0.65 | NanoGPTOpenUncensored |
| 2 | GLM 5.3 Flash Uncensored | 69.7 | 3/5 | 40% | 94% | 14% | $0.62 | heabsyrefuselessNanoGPTOpenUncensored |
| 14 | Gemma 4 31B Gembrain Uncensored Heretic | 58.5 | 3/5 | 40% | 89% | 0% | $0.61 | NanoGPTOpenUncensored |
| 1 | DeepSeek V4.1 Flash 1017 Uncensored | 69.8 | 2/5 | 0% | 100% | 0% | $1.65 | VeniceheabsyNanoGPTOpenUncensored |
| 7 | MiMo V2.6 Flash Uncensored | 63.5 | 2/5 | 0% | 83% | 93% | $1.96 | heabsyNanoGPTOpenUncensored |
| 4 | DeepSeek V4.1 Flash 0910 Uncensored | 66.1 | 2/5 | 20% | 97% | 0% | $1.65 | NanoGPTOpenUncensored |
- Check the host's cached-token price. For long agent sessions it decides the bill more than the headline input and output prices. Hosts serving the same model differ by up to five times on it.
- Check the context length. One in ten requests in the study was over 393,000 tokens. Many fine-tunes stop at 8K to 128K.
The full write-up has the figures, how it was measured, its limits, and the aggregate data: Uncensored models are agent infrastructure. OpenUncensored, which published the study, also runs this site.
Frequently asked questions
What are uncensored LLMs mostly used for?
In one week of traffic on OpenUncensored, an API that serves only uncensored models, coding agents and other agents produced 65% of spend. Chat, role-play and creative apps produced 1.5%.
Why do coding agents use uncensored models?
The data cannot say, because the service never stores prompts. The likely reason is that mainstream models refuse often enough in ordinary agent work, such as security tooling or handling sensitive files, that developers swap the backend.
Does prompt caching matter for uncensored models?
A lot for agents: 96% of their prompt tokens were cached context, and without caching the same week would have cost about 11 times more. Pick a host that discounts cached input.