Blog

Uncensored AI API for fewer refusals

Compare uncensored and abliterated models on one OpenAI-compatible API, with per-million-token pricing and 1M-context options.

2026-11-14

An uncensored AI API gives your application access to model variants that refuse less often than standard hosted models. Developers reach for them in roleplay, creative writing, security research, custom moderation work, and agent chains where a single refusal can break the whole run.

GeniusRelay exposes 43 uncensored and abliterated models from providers including DeepSeek, Z.ai, Xiaomi, Meta, and Community. Every model is available through one OpenAI-compatible endpoint, so you can move between a normal model and an uncensored edition by changing only the model name. All 599 models use the same key, the same base URL and the same billing balance.

What "uncensored" actually changes

Uncensored here means one of two things:

• DeAlign or derestricted fine-tunes. The model was trained to keep going on borderline prompts instead of stopping early.
• Abliterated checkpoints. The refusal direction is projected out of the weights, so the model is less likely to announce that it cannot help.

Neither guarantees zero refusals. A model can still be wrong, still hallucinate, and still stop when the context genuinely runs out. What changes is the shape of the failure: fewer dead ends, more usable first drafts, and far fewer retries in an automated pipeline.

Compare the options before you pick

Prices below are per million tokens from the model pages. Context is the maximum input window.

The full list lives on the uncensored models page, where you can sort by price and filter by context.

Which one should you start with

Start with a Heretic or Abliterated checkpoint when you mainly want fewer interruptions at a low price. Move to the DeAlign Uncensored edition when the first draft still stalls. If the workload is long (novel chapters, long agent memory), prefer the 1M-context options and budget for the input side, which dominates the bill on long prompts.

How to call it

Set the base URL to:

https://api.geniusrelaygpt.com/v1

Then send a normal chat completion:

curl https://api.geniusrelaygpt.com/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-uncensored",
    "messages": [{"role": "user", "content": "Continue the story."}],
    "temperature": 0.9
  }'

The same call from Python:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.geniusrelaygpt.com/v1",
    api_key="YOUR_KEY",
)

resp = client.chat.completions.create(
    model="deepseek-v4.1-flash-uncensored",
    messages=[{"role": "user", "content": "Continue the story."}],
    temperature=0.9,
)
print(resp.choices[0].message.content)

Streaming and tool calling use the standard OpenAI fields. There is a full API reference and a streaming guide.

Sizing the bill

Cost depends almost entirely on how much context you resend. A character chat that carries a 20K-token history sends 20K input tokens every turn. At $0.26/M input on GLM 5.3 Flash Uncensored, 1,000 turns costs roughly $5.20 in input alone. Trimming history or summarizing old turns usually saves more than switching to a cheaper model.

For one-shot generation — a chapter, a long prompt, a single rewrite — the uncensored edition at $0.429/M input is usually worth it over retrying a refusing model three times.

Use the pricing page calculator before committing to a volume, and set a per-key spend limit in the console so a looping script cannot drain the balance.

Who should use uncensored models

• Roleplay and character chat: fewer breaks in long sessions; pair with the roleplay models page.
• Creative writing: mature themes, violence, dialect and messy prose survive intact.
• Agent pipelines: a refusal is a failed step, so removing dead ends raises completion rates.
• Research and moderation: useful for building classifiers and safety datasets where you need raw behavior.

Limits worth knowing

• Uncensored is not correct. These checkpoints trade some instruction-following polish for freedom.
• Output price is where cost hides. Uncensored output rates run 3–7× the input rate; keep max_tokens tight.
• Thinking variants bill reasoning tokens. If you pick a :thinking model, the trace is charged as output.
• Rate limits still apply. Check error codes before load testing.

FAQ

Does uncensored mean unsafe?

No. It means the model refuses less often. You are still responsible for the terms of service and applicable law.

Can I switch models without changing my code?

Yes. Change the model field. The request and response format stays the same.

Is there a free trial?

Pricing is prepaid and usage-based. Start with a small balance and scale from there.

What is the difference between Heretic and Uncensored?

Heretic keeps the original mixed precision and removes alignment pressure from the weights. The Uncensored edition is a separate DeAlign fine-tune. They are different checkpoints, not two names for one model.

Can I use these with SillyTavern?

Yes. Point SillyTavern at the same base URL and pick a model. The setup guide walks through the four fields.

Will an uncensored model follow JSON schemas?

Most still support structured output, but expect more drift than a heavily aligned model. Validate the response before trusting it in an agent loop.

Next steps

Browse the full model list, run the comparison page against two candidates, and start with the cheapest one that passes your own prompts. The quickstart has copy-paste setup for every supported language.

A prompt pattern that uses the freedom

Uncensored models still respond to structure. Keep the format constraints and drop the hedging:

Write the next scene. Third person, past tense.
Do not summarize what came before. No preamble.
End mid-action.

Constraints do three things: they keep output on-brief, they bound max_tokens, and they make a refusal obvious — if the model opens with an apology instead of prose, the prompt is the problem, not the checkpoint.

Testing a new checkpoint

Run ten of your hardest prompts before committing to a model:

• Count refusals.
• Count how many outputs are usable without editing.
• Note the average output length; longer is more expensive.

Move to the next candidate if refusals stay above one in ten or the prose needs rewriting every time. Keep two checkpoints in config so a regression is a one-value rollback.

Related reading

• OpenAI-compatible API with 599 models
• DeepSeek API pricing, all variants
• Claude API alternative with OpenAI format