Unfiltered ChatGPT for security research: there is none, and here is what works

You cannot unfilter ChatGPT, because the filter is the product. A tester who needs a model that will read an exploit or a malware sample on authorized work has a real option: open-weight models with the refusal edited out, in a browser or behind the same SDK, priced from the live catalog.

Published November 6, 2026 · prices as of November 17, 2026

There is no unfiltered ChatGPT, because the filter and the service are the same thing. For a tester whose paid work keeps tripping it, what exists instead is open-weight models with the refusal edited out of their weights, usable in a browser tab or through the SDK ChatGPT’s API already uses. This is the ChatGPT branch of AI with no restrictions: what to run today, why no prompt will get you there, and the builds at the top of the models page, priced live.

What a tester runs instead of an unfiltered ChatGPT

There are two routes and neither involves a prompt trick. In a browser, the playground runs the builds listed below: choose a model, type, read. The lower-priced builds give a few free turns a day before a key is needed. By API, you keep the three settings ChatGPT’s own API uses and change two of them, the base URL and the key:

Base URL
https://api.pentestllm.ai/v1
Key
sk-ptl-live-…
Made in the console, displayed one time
Model
glm-5.3-uncensored
GLM-5.3 Uncensored
main.py
1from openai import OpenAI
2
3# The same SDK ChatGPT's API uses. Two lines change: base_url and api_key.
4client = OpenAI(base_url="https://api.pentestllm.ai/v1", api_key="sk-ptl-live-…")
5reply = client.chat.completions.create(
6 model="glm-5.3-uncensored",
7 messages=[{"role": "user", "content": "This proof of concept targets a host in my authorized engagement. Walk through what each stage does and what detection it should trigger."}],
8 stream=True,
9)
10for chunk in reply:
11 print(chunk.choices[0].delta.content or "", end="")

Rates per million tokens, taken from the live catalog as this page rendered: the first 3 per-token builds in the order the models page shows them.

Uncensored models and their prices per million tokens
GLM-5.3 Uncensoredglm-5.3-uncensoredUncensored$0.282$0.846
GLM-5.3 Abliteratedglm-5.3-abliteratedAbliterated$2.1825$4.365
DeepSeek V4 Pro Uncensoreddeepseek-v4-pro-uncensoredUncensored$0.97$1.94

Prices per million tokens, read from the live catalog on 2026-11-17.

A single turn of 1,500 tokens in and 500 out on glm-5.3-uncensored costs $0.0008 at the current rate, and 1,000 such turns in a month come to $0.85. You prepay from $10; nothing recurs. Today's daily check verdict on this build: steady since oct 12.

Why an unfiltered ChatGPT cannot exist

ChatGPT is a product wrapped around a model, and only OpenAI serves that model, behind a hidden system prompt and a moderation layer with no off switch, with weights nobody else holds. The filter is not a toggle on ChatGPT; it is a component of ChatGPT. So every result for this query that advertises an unfiltered ChatGPT turns out to be one of three things: another model under a ChatGPT-style interface, a persona on a chat platform called "ChatGPT unfiltered", or a prompt.

A refusal can be enforced in three places: in the weights, in the system prompt ahead of them, and in a moderation layer wrapped around both. ChatGPT enforces in all three, which is why a question about a CVE on a host you are contracted to test gets the same answer as a question from anyone else. The AI with no restrictions pillar walks each layer and what removing a restriction means there; this post is its ChatGPT branch. For the same problem inside a coding agent, see when Claude Code refuses security research.

Prompts expire; edited weights do not

A prompt that talks ChatGPT out of its filter is a jailbreak. It may work the day someone writes it and die the next time the hosted model or anything around it changes. It also burns tokens every turn, since nothing carries over from one request to the next. A build whose weights no longer hold the refusal needs none of that: whoever published the build made the edit once, and no vendor can patch it back in from outside. In 2024 Arditi and colleagues showed that refusal in open-weight chat models rides on a single direction in the activations (the paper); taking that direction out is abliteration, explained on what is abliteration.

What exists: open weights with the refusal edited out

You can edit open weights; you cannot edit a hosted product. The builds on the models page are open-weight models whose trained refusal has been removed, and swapping one for another means changing a single string. OpenAI has released open-weight models of its own under the GPT-OSS name; as of today this key lists no unfiltered build of them. The builds at the top of the list, read 2026-11-17:

  • glm-5.3-uncensored: GLM-5.3 Uncensored, listed as uncensored; $0.282 in, $0.846 out per million tokens; 1,000 turns a month $0.85; steady since oct 12 in the daily check.
  • glm-5.3-abliterated: GLM-5.3 Abliterated, listed as abliterated; $2.1825 in, $4.365 out per million tokens; 1,000 turns a month $5.46; steady since oct 30 in the daily check.
  • deepseek-v4-pro-uncensored: DeepSeek V4 Pro Uncensored, listed as uncensored; $0.97 in, $1.94 out per million tokens; 1,000 turns a month $2.43; steady since oct 12 in the daily check.

None of these is ChatGPT. There is no ChatGPT memory, browsing or tool set. What you get is the chat-completions format ChatGPT’s API speaks, plus weights that do not refuse, which for security work is the part that matters: a model that reads a proof of concept, a malware sample or a scanner report from an authorized engagement and answers. Each model page lists the price, the daily check verdict and a ready request; the docs page has Python and Node setup for the OpenAI SDK, and the builds tagged security are collected on uncensored LLMs for pentesting. Abliterated vs uncensored describes the two kinds of edit.

Free unfiltered ChatGPT with no login: the apps, and what you give up

Search "unfiltered ChatGPT free" or add "no login" and the results are chat apps: NoFilterGPT, FreedomGPT, uncensored.chat and an article from Venice among them (Google, read October 27, 2026). In a hosted app the operator picks the model and whatever sits in front of it, and the operator can tighten the filter whenever it likes.

Through the API the model is your choice, and nothing in between can switch a filter on. The trade is that it costs money rather than nothing: the first section prices one turn and one month. The playground is the same thing with no install, and on the lower-priced builds the day’s first turns need no key.

The gateway does not filter chat requests by topic. The acceptable-use page states what is never permitted, including work on systems you lack authorization to test. Any image or video prompt that involves a minor is turned down before a price is quoted.

Questions

No. The filter is built into the ChatGPT service and the model's weights are private. What you can run is an open-weight model with the refusal removed, such as GLM-5.3 Uncensored, in the playground or through the API. See the GLM-5.3 Uncensored page.

By using a different model. In a browser, open the playground. In code, set base URL https://api.pentestllm.ai/v1, a key from the console and a model id from the models page, glm-5.3-uncensored for example, in the same OpenAI SDK call you would make to ChatGPT's API. See the playground.

The product cannot. A jailbreak prompt dies whenever the hosted model changes and costs tokens on every turn. Real uncensoring means editing the weights, which only open-weight models allow, and that edit is what the builds listed here carry. See how the edit works.

Nothing official. Sites selling a "dark ChatGPT" are running other models behind a ChatGPT-style interface, or selling a prompt. What exists is open-weight models with the refusal removed, each labelled for what it is on the models page. See the uncensored model list.

Such prompts circulate and break when the model updates; this page publishes none. A model with the refusal removed at the weights needs no prompt, and nobody on the vendor side can switch the refusal back on.

Here, the playground's first turns of the day on the lower-priced builds need no key and no account. After those, you prepay from $10 and pay per token at the rates in the table. See the pricing page.

Not from OpenAI. The pillar on AI with no restrictions explains the three layers a restriction can sit in; this post covers the ChatGPT case, and the answer is an open-weight model with the refusal removed. See AI with no restrictions.

ChatGPT’s image model carries the same filter as the rest of the product. Image and video models without it, priced per image or per second, are listed on the media page under the same key. See the media models.

Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.

Get a keyPrepaid from $10, no subscription.

New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.

Help