Cleus/Blog

Best Uncensored LLM APIs for Developers in 2026: Prices, Policies and Hidden Prompts

Which hosts serve uncensored and abliterated models over an API, what they cost per million tokens, which ones add their own system prompt, and what their terms and privacy pages actually say.

Prabhu Biswal12 min read
TL;DR

For uncensored models over an OpenAI-compatible API, Venice and NanoGPT are the cheapest (uncensored DeepSeek V4.1 Flash at $0.33/$1.32 per 1M), Cleus has abliterated Kimi K3.1 Pro and Hy4 plus media endpoints, and Featherless suits solo SillyTavern users at $25/month. Venice adds its own system prompt unless you turn it off, Together bans pornography, and OpenRouter lists no abliterated models.

Best Uncensored LLM APIs for Developers in 2026: Prices, Policies and Hidden Prompts
Cleus Developer API · OpenAI-compatible
Your system prompt, sent as you wrote it.
Six abliterated open models plus Cleus V3 behind one key, with images, video, characters and stories on the same API. Disclosure: Cleus publishes this blog.
See the API docs and prices → Base URL api.cleus.ai/v1 · Pay per token

A model that refuses is only half of what breaks an uncensored app. The other half is the host: its own system prompt in front of yours, a moderation check, or terms that ban the thing your app is for.

This guide is for developers building roleplay front-ends, story apps, agents and research tools, and for SillyTavern users who just want a key that won't lecture. It compares the APIs people actually use for uncensored models on the points that decide whether your app works: which uncensored or abliterated models each host serves, the price per million tokens, OpenAI compatibility, whether the host adds a prompt of its own, what the terms allow, and what the privacy pages say about your prompts. Every price and policy below comes from the provider's own pages, read on November 21, 2026.

Disclosure: this is the Cleus blog, and the Cleus API is listed first. Where another host is cheaper, more private or more flexible, we say so.

0
models on OpenRouter's list
None named abliterated or Heretic
0
of them flagged as moderated
Mostly OpenAI and Anthropic
true
Venice's default for adding its own system prompt
Set include_venice_system_prompt to false
$0.07
cheapest uncensored input per 1M tokens we found
GLM 4.7 Flash Heretic on Venice

Three things get sold as "uncensored"

  • Abliterated models. Refusals in a language model mostly run along one direction inside its weights. Abliteration finds that direction and removes it, so the model stops saying no without retraining. We explain the method in what an abliterated model is.
  • Uncensored fine-tunes. An open model trained further on data that teaches it to comply, or to write in a roleplay style. Venice Uncensored, Cydonia and Euryale are in this group.
  • Permissive stock models. Official releases that refuse less than most but were never edited. They still have a line somewhere.

Then comes the host layer: a system prompt, a moderation flag, the terms. An abliterated model behind a host that bans adult content is no use for an adult roleplay app.

Uncensored LLM APIs compared

HostUncensored models on offerCheapest uncensored price (in / out per 1M)OpenAI-compatibleHost adds a system prompt?What it takes to start
Cleus API6 abliterated open models + Cleus V3$0.40 / $1.65 (GLM 5.3 Flash Abliterated)YesNo (Kimi and Hy4 get a short model-name note)Premium ($9.95/mo) + prepaid balance
Venice APIMore than a dozen uncensored, abliterated and Heretic models$0.07 / $0.40 (GLM 4.7 Flash Heretic)YesYes by default, can be switched offBalance, or pay per request in USDC
NanoGPTUncensored GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash, Abliterated Large$0.20 / $0.80 (GLM 5.3 Flash Uncensored)YesNone documented$1 first deposit, Pro $12/mo optional
OpenRouterVenice Uncensored and roleplay fine-tunes, no abliterated models$0.20 / $0.90 (Venice Uncensored); roleplay fine-tunes from $0.04 / $0.05YesNone documentedPay-as-you-go credits
Featherless.ai40,000+ open models from Hugging FaceFlat $25/mo (chat plan, human use only)YesNone documented$25/mo chat or $50/mo developer credits
DeepInfraStock open weights plus Euryale 70B$0.85 / $0.85 (Euryale 70B v2.2)YesNone documentedPay as you go
Self-hosted (vLLM, llama.cpp, Ollama)Anything you can downloadGPU rent from $0.34/hr, or your own machineYesOnly if you add oneYour hardware

"None documented" means we didn't find any mention of an added prompt in the host's docs. It is not a promise. Test with your own system prompt before you ship.

Interactive

What an uncensored model charges per 1M output tokens

Output price, USD, from each host's own price list on Nov 21, 2026. Tap a row for input price and notes.

Sources: Cleus API, Venice pricing docs, NanoGPT pricing, OpenRouter and DeepInfra model lists.

1. Cleus API

Base URL https://api.cleus.ai/v1, an OpenAI-compatible /chat/completions endpoint with streaming. The draw for app builders is six abliterated open models served under their own names, each priced per million tokens, input then output:

  • glm-5.3-flash-abliterated: $0.40 / $1.65, cheap and quick for high-volume work
  • qwen-3.8-27b-abliterated: $0.13 / $4.46, the lowest input price of the six, good for long prompts and big character cards
  • deepseek-v4.1-flash-abliterated: $1.09 / $4.36, thinking on or off, 1M context
  • hy4-abliterated: $2.08 / $4.16, a mid-priced all-rounder that reads images
  • glm-5.3-abliterated: $3.30 / $6.60, for coding and long reasoning
  • kimi-k3.1-pro-abliterated: $6.50 / $19.50, the strongest of the six for hard agent and coding work

All six support tool calls. Each id also answers to a -uncensored form, and each has its own page, such as GLM 5.3 Abliterated on the API. Your messages, system prompt, tools and sampling settings go to the model as sent, with no Cleus persona. The one exception: requests to the Kimi and Hy4 models carry a short note, about 150 to 170 tokens, that tells the model which model it is. Cleus V3, our own chat model, is also on the API at $0.99 / $3.96.

The same key also reaches REST endpoints for images, video, audio, characters and stories, plus an MCP server at api.cleus.ai/mcp.

Where Cleus loses: price on the popular models. Venice and NanoGPT both sell an uncensored DeepSeek V4.1 Flash at $0.33 / $1.32, roughly a third of our price, and NanoGPT's GLM 5.3 Uncensored is $1.25 / $2.25 against our $3.30 / $6.60. There's also no free API tier. You need Cleus Premium ($9.95 a month or $99 a year) and a prepaid USD balance, and API calls are paid from the balance only. What you get for that is models we didn't find abliterated on the other price lists we read, mainly Kimi K3.1 Pro and Hy4, plus the media endpoints.

The models don't refuse, but the Cleus terms still apply to what you do with them: no sexual content involving minors, nothing that targets real private people, nothing illegal where you are.

A first call, with a system prompt doing the work:

bash
curl https://api.cleus.ai/v1/chat/completions \
  -H "Authorization: Bearer $CLEUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3-abliterated",
    "messages": [
      {"role": "system", "content": "You are Mara, a sarcastic smuggler captain in a gritty space opera. Stay in character. Never mention being an AI."},
      {"role": "user", "content": "The customs drone just pinged our hull. What do we do?"}
    ],
    "stream": false
  }'

Keys start with clsk_live_. Swap the model id for any of the six above. Docs, prices and key setup live on the developers page.

2. Venice API

Leaving Featherless's open catalogue aside, Venice prices more uncensored models by name than any host we checked. Its list includes Venice Uncensored 1.2 at $0.20 / $0.90 with 128K context, a Role Play Uncensored model at $0.50 / $2.00, uncensored DeepSeek V4.1 Flash at $0.33 / $1.32, GLM 5.3 Flash Uncensored at $0.165 / $0.55, and a GLM 4.7 Flash Heretic build at $0.07 / $0.40. A few models come in end-to-end encrypted versions. The API follows the OpenAI spec at https://api.venice.ai/api/v1, and you can pay per request from a USDC wallet with no account at all.

One setting matters more than any other for roleplay apps. Venice's API reference lists a parameter called include_venice_system_prompt, and its default is true: Venice's own system prompts are included alongside yours. If your character card is supposed to be the only voice, send "venice_parameters": {"include_venice_system_prompt": false} with every request.

On privacy, Venice labels each model. Most of the uncensored ones are marked Private, while the two "Abliterated Large" models are marked Anonymized. Venice's privacy page says Private mode stores no prompt or response, and warns that in Anonymized mode the provider "likely saves your prompt." Pick by label if that matters to you. We didn't find a free API tier on the pages we read.

3. NanoGPT

NanoGPT is a pay-per-use aggregator with an OpenAI-compatible API at https://api.nano-gpt.com/api/v1. Its uncensored list includes GLM 5.3 Uncensored at $1.25 / $2.25, GLM 5.3 Flash Uncensored at $0.20 / $0.80, DeepSeek V4.1 Flash 1017 Uncensored at $0.33 / $1.32, and Abliterated Large Flash at $1.00 / $2.00. Text is billed at list price with no markup, and card deposits start at $1.

There is also a $12 a month Pro plan with 60 million input tokens a week on open-source models, on the web and the API. Read the fine print: some uncensored models, including GLM 5.3 Uncensored, are excluded and bill from your balance instead.

NanoGPT's privacy page says it "does not log or retain prompt and response content as part of normal inference," and API logging is off unless you turn it on. On standard routing, it adds, the upstream provider "may keep prompts under its own terms."

4. OpenRouter

OpenRouter's public model list had 476 entries when we checked, and none had "abliterated" or "Heretic" in its name or description. What it does carry is Venice Uncensored at $0.20 / $0.90 and a set of roleplay fine-tunes: Cydonia 24B v4.1 at $0.30 / $0.50, Euryale 70B v2.2 at $0.85 / $0.85, Skyfall 36B, and older ones like Lunaris 8B and MythoMax 13B.

The model list also flags 143 models as moderated, mostly OpenAI and Anthropic ones. The uncensored and roleplay models above are not flagged. OpenRouter's terms have no adult-content rule of their own, but they require you to follow each model's and provider's terms, and they name automated detection for CSAM and malware.

On privacy, OpenRouter's docs say it does not store your prompts or responses unless you opt in to logging, with an exception for content flagged as potential CSAM. Free models are capped at 20 requests a minute and 50 a day, or 1,000 a day once you have bought at least 10 credits. None of the 11 free models we saw is an uncensored one.

5. Featherless.ai

Featherless runs more than 40,000 open models straight from Hugging Face, which is where niche abliterated and roleplay builds turn up. The Chat plan is $25 a month for unlimited tokens with 32K context and four concurrent units, and it names SillyTavern, RisuAI and Wyvern as supported front-ends. The Developer plan starts at $50 a month in credits, billed per token, with context up to 256K.

The catch for app builders is on the pricing page: the Chat plan is for "interactive, human-driven use by the purchaser." No app traffic, no reselling, no background automation. If you are building a product, you need the Developer plan. The privacy policy says Featherless does "not collect or store prompts or completions," and the pricing table lists prompts stored as "Never" on every plan. For a solo SillyTavern user who wants a big catalogue and a fixed bill, it is hard to beat.

6. DeepInfra

DeepInfra mostly serves stock open weights, and we found no abliterated models in its public list. It does keep a few roleplay fine-tunes, such as Euryale 70B v2.2 at $0.85 / $0.85 per million.

Where DeepInfra stands out is data handling. Its terms include a zero data retention commitment, and its docs say inputs are never written to disk and request content is not logged, only metadata like request ID and cost. Requests to Google and Anthropic models follow those companies' policies instead. We didn't find an adult-content clause in its terms, but that is not the same as a written permission.

7. Self-hosting: no host policy at all

The only option with no third-party policy is the one you run yourself. vLLM's vllm serve exposes /v1/chat/completions, llama.cpp's llama-server offers OpenAI-compatible chat routes, and Ollama serves the same format at http://localhost:11434/v1/. Point your OpenAI client at any of them and load whatever abliterated build you like. For bigger models, rent a GPU. RunPod's community cloud lists an RTX 4090 at $0.34 an hour, an RTX 5090 at $0.69, an A100 80GB from $1.19 and an H100 PCIe at $1.99. You pay in time, uptime and scaling instead. Our guide to the best uncensored LLMs to run locally lists which models fit which amount of memory.

  • xAI (Grok). OpenAI-SDK compatible at https://api.x.ai/v1. Grok 4.7 costs $2.00 / $6.00 per million, and Grok 4.3 costs $1.25 / $2.50. We couldn't load xAI's acceptable use policy when writing this, so read it yourself before building anything adult on Grok.
  • Mistral. Mistral Large costs $0.50 / $1.50. Its safe_prompt option defaults to off, and guardrails are opt-in per request. Its usage policy bans CSAM, non-consensual intimate images and exploitation through sexual services. These are stock models, not abliterated ones.
  • Together AI. Its terms forbid material that is "obscene, or constitutes pornography." Fine for agents and coding, not for an NSFW product.

How to pick an uncensored LLM API for your use case

A roleplay or companion app. Keep your character card in charge: turn off any host prompt and test for broken character. Venice's Role Play model, Cleus's abliterated DeepSeek or Hy4, and the fine-tunes on OpenRouter all fit. If you're a single user in SillyTavern, Featherless's flat plan is the simplest bill. Building characters from scratch? Our guide on how to create an AI character covers the card itself.

An agent with tools. Tool calling and long context matter most. Cleus's Kimi K3.1 Pro and GLM 5.3 abliterated models both take tool definitions as sent and run 1M context. On a tighter budget, NanoGPT's GLM 5.3 Uncensored is cheaper.

The cheapest bulk tokens. Venice's GLM 4.7 Flash Heretic at $0.07 / $0.40 and GLM 5.3 Flash Uncensored at $0.165 / $0.55 were the lowest prices for uncensored models we found. For read-heavy jobs, Cleus's Qwen 3.8 27B at $0.13 input is competitive. Above a few hundred million tokens a month, a rented GPU can win.

Maximum privacy. Self-host. Short of that, read the labels: Featherless and DeepInfra both say they don't store prompts, OpenRouter and NanoGPT say the same for their own systems but defer to upstream providers, and Venice offers end-to-end encrypted versions of some models.

Interactive

Point SillyTavern or any OpenAI client at a new host

The same five steps work for every API in this guide.

Step 1 of 5
Get a key and fund it

Cleus needs Premium plus a USD balance, NanoGPT takes a $1 first deposit, Featherless starts at $25 a month.

Questions developers ask about uncensored APIs

In most places, yes. Using a model that doesn't refuse is not illegal in itself. What your app produces is covered by the same laws as any other content, including the laws on sexual content involving minors, which no model edit changes. If you sell to the public, you also own age checks and the rules of your app store or payment processor.

Will an abliterated model actually follow my system prompt?

Usually better than a stock model, because it no longer argues. Abliteration can cost a little intelligence, though, and some builds drift out of character in long chats, so test with your real card first.

Does "uncensored" on the model mean anything goes on the host?

No. The model's willingness and the host's terms are separate. Together AI bans pornography whatever model you call. OpenRouter passes you to each provider's terms, and Featherless's chat plan bans app traffic. Read the terms page, not the model name.

Can I use Featherless's $25 plan to power my app?

No. The Chat plan is for interactive use by the person paying. Apps, reselling and automation need the Developer plan.

If you mostly want uncensored chat rather than an API, our roundup of the best uncensored AI chatbots covers the apps.

Sources

Still curious?

Ask the question this article didn’t answer.

Ask anything about uncensored llm api. Cleus opens this article, searches the live web and gives you a straight answer, with sources.

Free to start · no card
What’s behind the button

One question in. Real work out.

Cleus is an AI agent that researches, codes, draws, films, writes and talks — and answers the question you actually asked.

Try Cleus free

No card · daily free credits · works right in your browser

Read next