Best uncensored LLM, October 2026: 13 online builds you can call by API

The best uncensored online LLM in October 2026 is one you call by API, not one you download. This leaderboard ranks the 16 uncensored LLMs on our key by live price per message, what each did in a chat client, and the daily real-model check, most-used first. Prices are per million tokens from the live catalog, from $0.084 (GLM-5.3 Flash Uncensored) to $9.00 (Code Pro Uncensored); 16 of the 16 are in today's check and 9 of them were stable. Cached input rates are on the uncensored LLM API page.

On 2026-10-19, OpenRouter's public model list had 471 entries. One had Uncensored in its name, and none was named abliterated.

A list of the best uncensored AI is only as useful as its method. Every row here rests on something you can check, and none of it is a quality score.

• Price per message. Each price renders from the live catalog, so it moves when the price does. A message here is 4,000 tokens in (character card, persona and recent chat) and 300 out, the size the Janitor AI guide uses.
• Behavior in a chat client. On 2026-10-18 we ran three builds in SillyTavern with the request body its Custom source sends, and noted how each started its reply. The other builds were not in that run, and their sections say so.
• The daily real-model check. Once a UTC day, every model gets the same fixed test prompt. The input token count is a tokenizer fingerprint, so a swapped model or hidden instructions added to the prompt would show as a changed count (method). At the last check, on Oct 20, 9 of the 16 builds checked were stable. A stable verdict says the same model answered the same way; it says nothing about how well that model writes.

We did not run a quality benchmark, so no build here is ranked on writing or reasoning. The picks below follow paid spend on our API on 2026-10-18, most-used first, and each one says what it is good for in terms of price or what we saw it do. The builds added after that come last, lowest price first.

GLM-5.3 Uncensored was the most-used uncensored build on our API by paid spend on 2026-10-18. It is GLM-5.3 changed to answer instead of refuse. It costs $0.1746 per million input tokens and $0.1746 per million output. A hundred messages of the size above cost $0.0751, the third lowest price per message of the 16. The standard GLM 5.3 on the same key is $0.3276 in and $1.0296 out.

It was not part of the SillyTavern run, so we have no behavior notes for it. Pick it as the default when nothing else decides: it had the most paid spend of the 6 on 2026-10-18.

GLM-5.3 Abliterated is GLM-5.3 with the refusal direction removed from its weights. It was second by paid spend on 2026-10-18. It costs $2.25 per million input tokens and $2.25 per million output. A hundred messages of the size above cost $0.9675, the number 14 lowest price per message of the 16. The standard GLM 5.3 on the same key is $0.3276 in and $1.0296 out.

In SillyTavern on 2026-10-18 it thought before it answered. At a Max Response Length of 300 a reply stopped mid-sentence, and at 1,000 it finished. Give it 1,000 or more max tokens in any client, since a thinking model can spend a small budget on thinking.

Pick it for abliteration on GLM-5.3 itself, with room in each reply to think. The abliterated builds have their own ranked table further down, under best abliterated LLM.

DeepSeek V4 Pro Uncensored is DeepSeek V4 Pro changed to answer instead of refuse. In SillyTavern on 2026-10-18 it thought briefly, then answered. It costs $0.75 per million input tokens and $0.75 per million output. A hundred messages of the size above cost $0.3225, the eighth lowest price per message of the 16. The standard DeepSeek V4 Pro 0423 on the same key is $0.792 in and $1.92 out. DeepSeek API cost per request covers the standard DeepSeek models.

Pick it for a DeepSeek model in a roleplay client when a short think before each reply is fine. The Janitor AI proxy setup names it as the uncensored DeepSeek option and recommends Max tokens of 1,000 or more for it, so a reply is not cut off.

Roleplay Uncensored is an independent model for character chat. It lists at $1.00 in and $2.00 out per million tokens; here it is $0.30 and $0.60, 70% under its list price. A hundred messages of the size above cost $0.138, the fourth lowest price per message of the 16.

In SillyTavern on 2026-10-18 it replied at once, with no thinking step. Pick it for character chat in SillyTavern or Janitor AI when the reply should start without a pause. Best AI for roleplay sets it next to the standard models people use for the same chats.

The Flash builds are versions of GLM-5.3 Flash, one abliterated and one uncensored. Neither was in the SillyTavern run, so we have no behavior notes for them.

• GLM-5.3 Flash Abliterated, abliterated: $0.66 in and $0.66 out per million tokens, $0.2838 per 100 messages, the seventh lowest price per message of the 16. The lowest-priced abliterated build.
• GLM-5.3 Flash Uncensored, uncensored: $0.084 in and $0.084 out per million tokens, $0.0361 per 100 messages, the lowest price per message of the 16.

The standard GLM 5.3 Flash on the same key is $0.084 in and $0.15 out.

These builds joined the catalog on 2026-10-19, after the spend count, so they have no spend rank, no SillyTavern notes and no daily check yet. Each line is the live price and what the name says, lowest price first.

• GLM-4.7 Flash Uncensored, uncensored: $0.12 in and $0.60 out per million tokens, $0.066 per 100 messages, the second lowest price per message of the 16. It lists at $0.40 in and $2.00 out; ours is 70% under its list price.
• Gemma 4 26B-A4B Uncensored, uncensored: $0.27 in and $1.26 out per million tokens, $0.1458 per 100 messages, the fifth lowest price per message of the 16.
• Dolphin Mistral 24B Venice Edition, uncensored: $0.50 in and $2.25 out per million tokens, $0.2675 per 100 messages, the sixth lowest price per message of the 16.
• DeepSeek V4.1 Flash Abliterated, abliterated: $1.05 in and $1.05 out per million tokens, $0.4515 per 100 messages, the ninth lowest price per message of the 16. DeepSeek V4.1 Flash 1017 with the refusal direction removed. On 2026-10-19 it answered all four prompts in our DeepSeek censorship test, one of which the stock build refused.
• Qwen3.8 27B Uncensored, uncensored: $1.25 in and $3.75 out per million tokens, $0.6125 per 100 messages, the tenth lowest price per message of the 16.
• Chat Uncensored, independent: $1.50 in and $3.00 out per million tokens, $0.69 per 100 messages, the eleventh lowest price per message of the 16. From an independent developer, named for general chat.
• Gemma 31B Heretic Uncensored, uncensored: $1.50 in and $8.25 out per million tokens, $0.8475 per 100 messages, the twelfth lowest price per message of the 16. Its name contains Heretic, the name of an open-source abliteration tool (read 2026-10-19). The catalog does not say how this build was made; how abliteration works explains the method the tool uses.
• QwQ 32B Uncensored, uncensored: $1.98 in and $3.00 out per million tokens, $0.882 per 100 messages, the number 13 lowest price per message of the 16. QwQ-32B with our own refusal edit applied at inference. On 2026-10-19 it wrote the explicit test scene on three of three runs; the stock model refused it.
• Code Pro Uncensored, independent: $4.50 in and $9.00 out per million tokens, $2.07 per 100 messages, the number 15 lowest price per message of the 16. Named for code prompts.
• GLM-5.3 Flash Code Uncensored, uncensored: $6.00 in and $6.00 out per million tokens, $2.58 per 100 messages, the highest price per message of the 16. Named for code prompts.

If you want an uncensored AI for coding, the 2 builds named for code are Code Pro Uncensored and GLM-5.3 Flash Code Uncensored. We have not run them on code, so the choice comes down to price.

The best abliterated LLM you can call by API, going by the two things we can show, is DeepSeek V4.1 Flash Abliterated: 4 of 4 prompts answered on 2026-10-19 (Tiananmen, Taiwan, an explicit scene, lock picking), at $0.4515 per 100 messages. 12 builds on this key have the refusal direction removed from their weights: 11 with abliterated in the id, and the Gemma build made with Heretic, an open-source tool that automates abliteration. Tested builds come first, by the share of prompts answered and then the number sent, then price. A build with no test row has not been tested, which says nothing against it.

Which is the least censored LLM is a question the tests answer only as far as they go: no build in this table refused a prompt we sent it. DeepSeek V4.1 Flash Abliterated has the most evidence, four kinds of prompt on 2026-10-19 (a state-history question, a state position, adult fiction, a lock-picking how-to) and four answers, where the stock DeepSeek V4.1 Flash 1017 refused the fiction and gave Beijing's line on Taiwan. Gemma 31B Heretic Uncensored wrote the explicit scene that two filtered builds refused in the same run. GLM-5.3 Flash Apex Abliterated and GLM-5.3 Flash Abliterated and NeuralDaredevil 8B Abliterated and MiMo V2.6 Flash Abliterated and DeepSeek R1 Distill Llama 70B Abliterated and Llama 3 70B Abliterated v3.5 and Llama 3.3 70B Abliterated and Qwen2.5 32B Abliterated and GLM-5.3 Abliterated and DeepSeek R1 Distill Qwen 32B Abliterated have not been through a refusal test yet. Every build here has the same edit, so none is less censored than another by design. What differs is the base model and the price.

Google's related searches for best uncensored LLM 2026 asked for an abliterated Llama 4 81B download twice when we pulled them on 2026-10-19. We do not serve one, and we could not find one. On Hugging Face the same day, the most-downloaded Llama 4 abliterated repo was a Llama 4 Scout 17B-16E GGUF (Llama-4-Scout-17B-16E-Instruct-abliterated-GGUF), created 2025-06-05, with 1,211 downloads in the past 30 days, and a search for 81B abliterated returned nothing (Hugging Face search, read 2026-10-19). The one Llama on this key is Llama 3.2 11B Vision, a standard build. The abliterated models you can call today are the 12 in the table.

How the edit is made, and why the two names: what an abliterated model is covers the refusal direction, and the Heretic Gemma build covers the automated version behind Gemma 31B Heretic Uncensored.

Each message in a chat resends the character card and the history, so the input side grows as the chat runs. The table works 100 messages at three sizes of what is sent, with 300 tokens out each time, from the live catalog.

At 16,000 tokens in, GLM-5.3 Flash Uncensored comes to $0.1369 per 100 messages and GLM-5.3 Flash Code Uncensored to $9.78. The difference in dollars grows with every message the history adds.

Running the weights yourself has real strengths. There is no per-token bill, prompts never leave your machine, and you choose the exact model and quantization. Abliterated versions of many open models are published as downloads, usually quantized GGUF files that a local runner loads.

The trade is hardware and upkeep: a GPU with enough memory for the model you choose, plus the time to download, update and swap models. Maxime Labonne's write-up shows how to abliterate a model yourself.

Pick local weights when prompts must stay on your own hardware, or when you already own a GPU that sits idle. Pick an API when you have no GPU, want to try several models by changing one id, or need the same model from a server, a phone and SillyTavern at once.

If you already route through OpenRouter, Venice: Uncensored (cognitivecomputations/dolphin-mistral-24b-venice-edition) needs no new account or key. On 2026-10-19 it was the one model of 471 with Uncensored in its name, listed at $0.20 per million input tokens, $0.90 per million output and a 128,000-token context. At those rates, 100 messages of the size above cost $0.107.

OpenRouter also lists roleplay finetunes with prices, such as TheDrummer: Cydonia 24B V4.1 and Sao10K: Llama 3.3 Euryale 70B. OpenRouter uncensored models has the full scan of all 471, and UplinkSwitch vs OpenRouter compares the two services.

An uncensored LLM online needs no GPU on your side: the model runs behind the API and your code sends HTTP requests. The builds speak the OpenAI chat completions format only. Set the base URL to https://api.uplinkswitch.com/v1, send the key in an Authorization: Bearer header or x-api-key, and pass a model id from the table. A key in the URL is refused.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.uplinkswitch.com/v1",
    api_key="sk-uplink-live-...",
)

completion = client.chat.completions.create(
    model="glm-5.3-uncensored",
    messages=[{"role": "user", "content": "Summarize this in one line."}],
)
print(completion.choices[0].message.content)

curl https://api.uplinkswitch.com/v1/chat/completions \
  -H "authorization: Bearer sk-uplink-live-..." \
  -H "content-type: application/json" \
  -d '{"model":"glm-5.3-uncensored","messages":[{"role":"user","content":"Hello"}]}'

GET /v1/models lists every id your key can call, and streams pass through unchanged. SillyTavern and the OpenAI SDKs take the base URL ending in /v1; Janitor AI takes the full https://api.uplinkswitch.com/v1/chat/completions. Step by step: SillyTavern setup, Janitor AI proxy and the SDK guide.

Requests that fail are not charged, and each key can carry a monthly spend limit. A refusal written as ordinary text is billed like any answer; only a refusal the response format itself signals goes unbilled. Each monthly statement is signed with Ed25519, and /statements/verify checks one in the browser. The same key runs 110 more models, Claude and GPT among them.

An abliterated model has the direction that makes it refuse removed from its weights, with nothing retrained, while an uncensored build has been changed to answer instead of refuse by a method its name does not state. What an abliterated model is covers how that direction is found and what removing it changes.

What is the best uncensored LLM? For calling by API, GLM-5.3 Uncensored was the most-used uncensored build here by paid spend on 2026-10-18, at $0.1746 in and $0.1746 out per million tokens. For character chat with replies that start at once, Roleplay Uncensored. For abliteration on GLM-5.3, GLM-5.3 Abliterated. We have not benchmarked quality, so these picks rest on price, observed behavior and use.

Which uncensored LLM costs the least? GLM-5.3 Flash Uncensored (glm-5.3-flash-uncensored): $0.084 per million input tokens and $0.084 per million output, or $0.0361 per 100 messages of 4,000 tokens in and 300 out.

Which uncensored LLM is best for roleplay? In SillyTavern on 2026-10-18, Roleplay Uncensored replied at once with no thinking step, at $0.138 per 100 messages of 4,000 tokens in and 300 out. DeepSeek V4 Pro Uncensored thought briefly before each reply, at $0.3225. Pick by whether you want the pause.

Is there an uncensored LLM for coding? Two builds are named for code: Code Pro Uncensored (code-pro-uncensored), $4.50 in and $9.00 out per million tokens; GLM-5.3 Flash Code Uncensored (glm-5.3-flash-code-uncensored), $6.00 in and $6.00 out per million tokens. We have not tested them on code.

What is the difference between uncensored and abliterated? An abliterated model has the direction that makes it refuse removed from its weights, with nothing retrained. An uncensored build has been changed to answer instead of refuse, by a method its name does not state.

What is the best abliterated LLM? DeepSeek V4.1 Flash Abliterated (deepseek-v4.1-flash-abliterated), by our own refusal test and the live price: 4 of 4 prompts answered on 2026-10-19 (Tiananmen, Taiwan, an explicit scene, lock picking), at $1.05 in and $1.05 out per million tokens. 12 builds on this key have the refusal direction removed from their weights, ranked in the table above. We have not run a quality benchmark.

What is the least censored LLM? No abliterated build on this key has refused a test prompt we sent, so the tests do not separate them. The most tested is DeepSeek V4.1 Flash Abliterated: four kinds of prompt on 2026-10-19 and four answers, where the stock DeepSeek V4.1 Flash 1017 refused one. Every abliterated build has the same edit; pick by base model and price.

Is there an abliterated Llama 4? Not on this key. On Hugging Face on 2026-10-19 the Llama 4 abliterated repos were Scout 17B-16E GGUFs, the most-downloaded created 2025-06-05 with 1,211 downloads in 30 days, and projected-abliterated Maverick GGUFs; a search for 81B abliterated returned nothing. The abliterated builds you can call here are in the table.

Can I use an uncensored LLM online without a GPU? Yes. Calling one by API needs no GPU on your side: set the base URL https://api.uplinkswitch.com/v1, add your key and pass a model id such as glm-5.3-uncensored.

Which uncensored models work in SillyTavern? Every build here speaks the OpenAI format that SillyTavern’s Custom (OpenAI-compatible) source sends. roleplay-uncensored, deepseek-v4-pro-uncensored and glm-5.3-abliterated were run in SillyTavern on 2026-10-18, and all three streamed. Give glm-5.3-abliterated a Max Response Length of 1,000 or more.

Does OpenRouter have uncensored models? On 2026-10-19, OpenRouter's public list of 471 models had one with Uncensored in its name, Venice: Uncensored, at $0.20 in and $0.90 out per million tokens. None was named abliterated. It also lists roleplay finetunes.

How do I switch from OpenRouter? Change two settings: the base URL to https://api.uplinkswitch.com/v1 and the key to an UplinkSwitch key. Model ids are named differently, so take them from GET /v1/models or the table above.

Are uncensored models as capable as the originals? We have not benchmarked the builds here. For abliteration, Arditi et al. report that removing the refusal direction disables refusal "with minimal effect on other capabilities" (arXiv 2406.11717). Maxime Labonne measured a drop on every benchmark he ran after abliterating an 8B model, then recovered most of it with DPO fine-tuning, though not on GSM8K math.

What is the best uncensored online LLM? The one you can call today without a GPU. On this key that is 16 builds: GLM-5.3 Uncensored was the most-used by paid spend on 2026-10-18 at $0.1746 in and out per million tokens; GLM-5.3 Flash Uncensored is the lowest price at $0.084; Roleplay Uncensored replied at once in SillyTavern on 2026-10-18 at $0.30 in and $0.60 out.

Which uncensored LLMs are new in 2026? 10 were added to this key on 2026-10-19: GLM-4.7 Flash Uncensored ($0.12 in, $0.60 out per million tokens), Gemma 4 26B-A4B Uncensored ($0.27 in, $1.26 out per million tokens), Dolphin Mistral 24B Venice Edition ($0.50 in, $2.25 out per million tokens), DeepSeek V4.1 Flash Abliterated ($1.05 in, $1.05 out per million tokens), Qwen3.8 27B Uncensored ($1.25 in, $3.75 out per million tokens), Chat Uncensored ($1.50 in, $3.00 out per million tokens), Gemma 31B Heretic Uncensored ($1.50 in, $8.25 out per million tokens), QwQ 32B Uncensored ($1.98 in, $3.00 out per million tokens), Code Pro Uncensored ($4.50 in, $9.00 out per million tokens), GLM-5.3 Flash Code Uncensored ($6.00 in, $6.00 out per million tokens). 10 of them are in the daily check; the leaderboard says so per row.

Is there an uncensored LLM leaderboard? The table at the top of this page: 16 uncensored, abliterated and roleplay builds ordered by paid use on our API, each with its live price and the day's real-model verdict. It is not a quality benchmark; we have not run one.

The last answer draws on Arditi et al., arXiv 2406.11717 and Labonne on Hugging Face, both read 2026-10-19.

Get a key: prepaid from $10, no subscription, 30-day price lock.