The uncensored AI API for developers

One endpoint for 8 uncensored models and 178 frontier models. Change the base URL, keep the OpenAI chat format, and pay only for the tokens you use.

Last updated: 2026-10-23

What you get

This is a narrow, developer-first gateway rather than a consumer chat app. The product is the endpoint: POST https://abliteratly.beer/v1/chat/completions. It accepts the OpenAI chat-completions shape, including streaming, tool calls, JSON mode and vision where the selected model supports them. You can keep an existing OpenAI SDK and change two values: the base URL and the model ID.

Endpoint

https://abliteratly.beer/v1

Models

8 uncensored / 178 frontier

Auth

Bearer key from the dashboard

Payment

Crypto via NOWPayments; minimum $20

Account

Account ID and password, no email

Prompt storage

Prompt and response bodies are not stored by us

First request

Create a key in the dashboard, then send a chat completion. The example uses the first uncensored model in the live catalog. Any enabled model ID from the directory works.

# change base_url and key. nothing else.
curl https://abliteratly.beer/v1/chat/completions \
  -H "Authorization: Bearer $ABLITERATLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "magnum-v4-72b",
    "messages": [{"role": "user", "content": "Write a scene with moral ambiguity."}]
  }'

The same request with the OpenAI Python SDK:

from openai import OpenAI

client = OpenAI(
    base_url="https://abliteratly.beer/v1",
    api_key="sk-ub-..."
)

r = client.chat.completions.create(
    model="magnum-v4-72b",
    messages=[{"role": "user", "content": "Write a scene with moral ambiguity."}]
)
print(r.choices[0].message.content)

For a guided setup, use the quickstart or the OpenAI SDK notes. OpenCode users can run the one-command installer.

Uncensored models in the catalog

These are the uncensored and abliterated models currently enabled. Prices are USD per 1M tokens and update with the catalog. A model page has the context window, tool support and a model-specific example.

What "uncensored" actually means

Most mainstream chat models are trained and served with refusal behaviour. An uncensored API does not remove the need for judgment; it changes which model behaviour you can choose. There are three common approaches:

• Abliterated models. The refusal direction is estimated in the model's weights and reduced or removed. This is a weight-level intervention, not a system prompt. It can change tone and coherence, so evaluate each model rather than assuming the label is enough.
• Fine-tunes. A model is trained on a dataset with different refusal behaviour. The result can be more natural than abliteration, but quality depends heavily on the dataset and base model.
• Prompt steering. A system prompt asks a normal model to comply. It is the cheapest option and the least reliable; the underlying refusal behaviour is still there.

The abliterated models guide explains the trade-offs in more detail. No model is guaranteed to answer every prompt, and every model remains subject to the acceptable use policy.

Why use an API instead of a chat app?

A chat app is built for one person typing. An API is built for software: repeatable model IDs, usage accounting, rate limits, streaming, tool calls and a bill you can attribute. If you are building an agent, a fiction tool, a research workflow or an internal assistant, the API is the part you can actually ship.

Limits you should know

The gateway speaks chat completions, not the full OpenAI platform. Responses, embeddings, audio, image generation and fine-tuning endpoints are not offered. Model availability changes with the upstream catalog. If a model is disabled, requests return 404 model_not_found. Use the model directory as the source of truth, not a blog post.

Latency and refusal behaviour vary by model and route. For production, add timeouts, retry 502/503 responses with backoff, and handle 429 with Retry-After. Do not assume a model will behave the same way after an upstream update.

Pricing and payment

There is no subscription. You buy credit with crypto, and the USD value actually received is applied after confirmation. Credits do not expire. Uncensored models carry the margin; frontier models are passed through at the upstream list price. See the pricing page for the live rate sheet and a request-cost calculator.

How it compares

If breadth is the priority, read the OpenRouter alternatives comparison. If you want a consumer chat product rather than an endpoint, see the Venice AI alternative notes.

Common use cases

• Creative writing and roleplay. Choose a model for voice and coherence, not just refusal rate. Use a system prompt for format and a separate model for editing.
• Coding agents. Use a frontier model for tool-heavy code work, and an uncensored model where a policy refusal would break the flow. See coding agents.
• Research and red-teaming. Keep prompts and outputs in your own system; the API does not store message bodies. Follow the acceptable use policy.
• Internal tools. One key and one balance can cover multiple model IDs, with usage metadata for attribution.

Get started

Create an account, make a key, and send the first request. There is no email step. Save the account ID: it is the only login and there is no email recovery.

Questions