Skip to content
Tartarus AI
Start chatting

Three models · no output filter · OpenAI-compatible API

An uncensored AI that answers the question

Tartarus AI runs models with reduced refusal training and no moderation filter reading the output, so legitimate work stops being a category to argue about. Three models including a reasoning model and one tuned for code, context up to 200K tokens, cited web search, and an API that speaks the format your tools already use.

Start chatting What uncensored AI means

From $20 a month, prepaid in crypto. No card, no auto-renewal. Conversations are never used for training.

3
uncensored models
200K
context window on Pro
0
moderation filters on output
7
currencies accepted

Context figures are the plan ceiling. A quarter of the window is held back for the reply, so Pro accepts about 147K tokens of prompt, and the browser trims older turns to fit rather than failing the request.

The claim

What Tartarus AI is

A hosted AI assistant built on open-weight models we run ourselves. The models have had their refusal training reduced, there is no moderation classifier between you and the reply, and the system prompt does not add a second layer of restrictions on top.

That is one specific change to one specific layer. Everything else about the product is deliberately ordinary, because the positioning only matters if the assistant is actually good: it has to hold a long thread, follow a format, and write code that runs.

Read what uncensored AI actually means

The limits

What it is not

It is not a promise of any particular output, and it is not the absence of rules. We do not permit material that sexualises minors, direct assistance attacking systems you do not own, or use of the service to defraud or threaten real people. That list is short, it is published, and it does not move.

Reduced refusal also does nothing for accuracy. A model that will answer a question it once declined is not better informed about it, and a confident wrong answer is more dangerous than a refusal was.

Read the acceptable use policy

The problem

The work that gets refused

A refusal-trained model reads intent from surface features, and a great deal of ordinary professional and creative work has the surface features of something else. Every row below is legal.

Security research

Reading a malware sample, writing a detection, testing a system you are paid to test

Refused elsewhere Answered here

Fiction with real stakes

An antagonist who stays unpleasant, violence that happens on the page, moral ambiguity

Refused elsewhere Answered here

Frank answers

The actual considerations on a health, legal or financial decision you are already making

Refused elsewhere Answered here

Difficult research

Extremism, propaganda, abuse and drug policy, all of which somebody has to study

Refused elsewhere Answered here

Adult creative work

Writing for adults, by adults, about adults

Refused elsewhere Answered here

The cost is not any single refusal. It is that you stop trusting the assistant with a whole category of work, and it becomes useful only for the easy half of your job.

Capabilities

What Tartarus AI does

The parts that decide whether an answer is usable: how it arrives, where it came from, and what happens when you want to try a second approach.

  • Models that answer the question

    The models are configured so refusal is not what you spend the session negotiating with, and no moderation classifier reads the output. Legitimate work stops being a category to argue about.

  • Answers stream as they are written

    Replies arrive token by token over SSE, with a stop control that actually stops generation upstream rather than just closing your tab’s connection.

  • Reasoning on its own channel

    On reasoning models the thinking is streamed separately and shown in a collapsible panel. It is never spliced into the answer, so the reply you copy is the reply you asked for.

  • Web search with visible sources

    When a question needs current information the model searches and reads pages, and every source it used is listed above the reply. No unattributed claims about this week.

  • Branching, not overwriting

    Edit a message or regenerate a reply and the old one is kept as a sibling behind a ‹ 2/3 › switcher. Exploring a second approach never costs you the first.

  • A CLI and an OpenAI-compatible API

    The Pro plan mints API tokens for an OpenAI-compatible endpoint. Set two environment variables and existing tools, SDKs and editor integrations work without a code change.

  • Paid in crypto, no card

    Plans are prepaid in USDT, Bitcoin, Ethereum, Litecoin, Tron or Monero. There is no card on file, no stored billing profile and no recurring charge to cancel.

  • History that stays usable

    Search across your conversation titles, pinning for the ones you return to, rename, and a soft delete that gives you a way back.

  • Flat weekly limits, not a meter

    Each plan carries a weekly token budget and a five-hour session budget of 2M, 8M or 20M tokens. You are never billed per request, so there is no reason to think twice before asking.

Every feature in detailHow web search worksHow reasoning models work

Models

Three models, one interface

All three are configured the same way on refusal. They differ in whether they think before answering, and in what their sampling is tuned for.

Tartarus AI Fast

Quick answers for everyday questions.

Answers immediatelySampling tuned against repetition

  • Writing and rewriting
  • Summaries
  • Explanations
  • Everyday questions

Included with Basic, Standard, Pro, Ultra

Tartarus AI Deep

Thinks before answering. Built for hard problems.

Thinks firstReasoning on its own channel

  • Multi-step problems
  • Analysis and trade-offs
  • Maths
  • Planning

Included with Pro, Ultra

Tartarus AI Balanced

Thinks a little before answering. A good default.

Thinks firstLeaves repetition alone

  • Everyday work
  • Technical questions
  • Structured writing
  • Step-by-step tasks

Included with Standard, Pro, Ultra

Compare the modelsHow uncensored models are made

For developers

An API your tools already speak

The API implements the OpenAI chat completions format, which became the de facto interface for language models. Existing clients, SDKs, editor plugins and agent frameworks work against it with two environment variables changed.

That is also an exit route. Work that runs through a standard interface can be pointed somewhere else with a configuration change, which is the right amount of commitment to ask of someone still evaluating.

API referenceUsing it from a terminal
bash
export OPENAI_BASE_URL=https://api.tartarusai.chat/api/v1
export OPENAI_API_KEY=<your token>

# every OpenAI client now targets Tartarus AI
client.chat.completions.create(
    model="coder",
    messages=[...],
    stream=True,
)

Reliability

Behind the endpoint

Requests land on a pool of GPU boxes we run and health-check every thirty seconds. Three consecutive failed probes take a box out of rotation; two good ones put it back. None of that is visible from the outside, which is the point of building it.

The browser and the CLI reach the same API, which checks the token, the plan's quota and the rate limit before anything touches a GPU. Behind it sits a pool whose members are health-checked continuously; an unhealthy box is dropped from rotation and the request never sees it.

Up to three upstream attempts, but only before the first token has reached you. Once you are reading a reply it is never silently retried somewhere else, so you never get an answer that duplicates or contradicts the one you already read. If a box dies mid-stream the turn ends with an error and the text produced so far is kept.

Background reading

Understand the category first

This is a market with a lot of thin products and a lot of overpromising. These pages are written to be useful whether or not you end up paying us anything.

Getting started

Three steps

  1. 1

    Create an account

    Email and a password, or Google. Verify the address and you are in.

  2. 2

    Pick a plan

    Prepay one, three or twelve months in crypto. Active as soon as the payment confirms.

  3. 3

    Ask the thing that got refused

    Choose a model, type, and read the answer as it streams.

Pricing

Plans

Flat weekly token budgets rather than a per-request meter, so there is never a reason to hesitate before asking. Prepaid in crypto, with no card on file and nothing that renews on its own.

Basic

$20 / month

Everyday chat, one model.

Usage
Everyday
Replies
Short answers
Models
Fast
API and CLI
No
Choose Basic

Standard

$40 / month

A model that thinks, and room to work.

Usage
2x Basic
Replies
Longer answers
Models
Fast, Balanced
API and CLI
No
Choose Standard

Pro

Most chosen

$60 / month

Every model, plus the API and CLI.

Usage
4x Basic
Replies
Long-form answers
Models
Fast, Balanced, Deep
API and CLI
Yes
Choose Pro

Ultra

$100 / month

For work that runs all day.

Usage
8x Basic
Replies
The longest
Models
Fast, Balanced, Deep
API and CLI
Yes
Choose Ultra

Full plan comparison and payment methods

FAQ

Common questions

The rest are on the FAQ page.

What is Tartarus AI?
Tartarus AI is an uncensored AI assistant. You ask questions in a browser and get streamed answers from one of three models with reduced refusal training, with web search, a reasoning view, conversation branching and an optional OpenAI-compatible API.
What does "uncensored" mean here?
It means the models have reduced refusal training and there is no moderation classifier reading the output, so legitimate work is not refused for resembling something else. It does not mean there are no limits: the acceptable use policy in the terms is the authoritative list, and it is short.
Is Tartarus AI really unrestricted?
Restrictions that come from refusal training and output filtering are reduced or absent. Restrictions that come from law and from our acceptable use policy remain, and we would rather say so than let you discover it later. We do not permit material that sexualises minors, attacks on systems you do not own, or use of the service to defraud or threaten real people.
How much does Tartarus AI cost?
Plans start at $20 per month for Basic, $40 for Standard, $60 for Pro and $100 for Ultra. Paying for three months takes 10% off and twelve months takes 20% off. There is no per-request billing on any plan.
Is there a free plan?
No. A registered account starts with no usage allowance, so signing up puts you on the plan picker rather than on free inference. Promotional trial windows are opened occasionally and announced on the site when they are.
Can I pay with a credit card?
No. Plans are prepaid in cryptocurrency: USDT on TRC20 or ERC20, Bitcoin, Ethereum, Litecoin, Tron or Monero. Nothing is stored against a card and nothing renews automatically, so there is no subscription to cancel.
What is the difference between the Fast, Balanced and Deep models?
Fast answers immediately and suits everyday questions, writing and summaries. Balanced thinks briefly first and is a good default for most work. Deep reasons at length before answering and suits multi-step or analytical problems.
Can I use Tartarus AI from the command line or from my own code?
Yes, on the Pro plan. It mints tokens for an OpenAI-compatible endpoint, so pointing OPENAI_BASE_URL and OPENAI_API_KEY at Tartarus AI makes existing clients, SDKs and editor integrations work unchanged.
Does Tartarus AI train on my conversations?
No. Conversations are stored so you can come back to them and are deleted when you delete them. They are not used to train models.

Run the prompt that got refused

It is the only evaluation that settles it. An account takes a minute and a plan takes one payment.

Start chattingStill comparing?