170 uncensored LLMs sit behind one key here today, and each of them also answers in the browser at /chat, three turns a day free with no account. This page puts them in order. A model earns a row by passing the real-model check that runs every day. The rows are then sorted by what a million output tokens cost. When UGI has scored the same weights, its figure is printed next to the row.
147 of 170 models passed the check that ran at 2026-11-16 02:56 UTC. Prices come from the November 16, 2026 catalog. Nothing here is typed in: every number is read when the page renders, and the page re-renders at least once an hour.
The November 2026 board
GLM-5.3 Flash Uncensored tops the November 2026 board at $0.3685 per million output tokens, the lowest price of any model that passed. 8 of the 147 rows charge less than one dollar for a million output tokens.
| # | Model | Build | Out $/M | Window | Verdict |
|---|---|---|---|---|---|
| 1 | GLM-5.3 Flash Uncensored | Uncensored | $0.3685 | — | Steady since Oct 12 |
| 2 | MythoMax L2 13B | Roleplay | $0.404 | 4K | Steady since Oct 20 |
| 3 | Manta Flash 1.0 | Roleplay | $0.56 | 16K | Answered the check |
| 4 | Manta Mini 1.0 | Roleplay | $0.56 | 8K | Answered the check |
| 5 | Cydonia 24B v4.3 | Roleplay | $0.60 | 131K | Steady since Oct 23 |
| 6 | GLM-5.3 Uncensored | Uncensored | $0.846 | — | Steady since Oct 12 |
| 7 | Gemma 4 12B Starry Ocean | Roleplay | $0.875 | 131K | Steady since Oct 23 |
| 8 | Gemma 4 12B StationKeeper | Roleplay | $0.875 | 131K | Steady since Oct 23 |
| 9 | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | Abliterated | $1.0389 | 262K | Steady since Oct 23 |
| 10 | Gemma 4 31B Split-Untied | Roleplay | $1.05 | 262K | Steady since Oct 23 |
What to pick for the job
Price sets the order above. The table below reads the same rows by purpose. For each job it takes the first row of the matching build kind that passed today, or the UGI figure quoted beside it.
| Job | Model | Reason, from today’s rows |
|---|---|---|
| Spend the least per answer | GLM-5.3 Flash Uncensored | The lowest output price of any row today, $0.3685 per million tokens |
| Weights with the refusal direction cut out | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | The abliterated build that passed at the lowest price, $1.0389 per million output tokens |
| Character chat through a card client | MythoMax L2 13B | The roleplay build that passed at the lowest price, $0.404 per million output tokens |
| Offensive security work | Gemma 4 26B A4B Cybersecurity | The security build that passed at the lowest price, $1.3376 per million output tokens |
| Feed it a whole codebase or a book at once | GLM-5.3 Flash Cybersecurity | Biggest context window of any row that passed: 1,048,576 tokens |
| The build that has stayed itself longest | GLM-5.3 Flash Uncensored | Steady since Oct 12: no other row has held steady as long |
| Top UGI score among the rows that passed | Cydonia 24B v4.3 | UGI 44.52 and W/10 7 on the UGI Leaderboard, tested December 16, 2025 |
| Most willing, by UGI | Dolphin Mistral 24B Venice Edition | W/10 7.8: the top willingness figure of the UGI rows that passed today (tested October 4, 2025) |
Read the build column before the price column. Abliterated means the refusal direction was cut out of the weights after training. Uncensored means the model was fine-tuned until it answers. A roleplay tune learned on long character chat, and tends to be a smaller, lower-priced model. The first two are compared in Abliterated vs uncensored.
Abliterated builds that passed
The catalog carries 38 abliterated builds and 36 of them passed the Nov 16 check. Same order as above: passed, then the lowest output price first. Builds named Heretic, derestricted or unshackled count too, since the catalog tags every build with the refusal cut from its weights as abliterated.
| # | Model | Out $/M | Window | Verdict |
|---|---|---|---|---|
| 1 | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | $1.0389 | 262K | Steady since Oct 23 |
| 2 | Qwen3.5 27B Blossom V6.4 Derestricted | $1.071 | 262K | Steady since Oct 23 |
| 3 | Qwen3.5 27B BlueStar v3 Derestricted | $1.071 | 262K | Steady since Oct 23 |
| 4 | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted Lite | $1.071 | 262K | Steady since Oct 30 |
| 5 | Qwen3.5 27B Queen Derestricted | $1.071 | 262K | Steady since Oct 23 |
| 6 | DeepSeek V4 Pro Qwen3.5 9B Uncensored Heretic | $1.12 | 33K | Steady since Nov 8 |
| 7 | Gemma 4 31B Gembrain Heretic | $1.224 | 262K | Steady since Oct 23 |
| 8 | Gemma 4 31B SDFT Heretic RP | $1.224 | 262K | Steady since Oct 23 |
Any abliterated build not shown here is in the full table further down, with the same columns.
Uncensored LLMs you can use online, right now
Nothing on this board is downloaded. Open /chat and the model answers in the browser, with a system prompt and temperature in Settings and nothing to install; a phone, a laptop and a server all reach the same id. The playground runs any row on a key in a browser tab, and the client docs show where the base URL and id go in Open WebUI, SillyTavern, Janitor AI or the OpenAI SDK.
By price, the online pick is the first row, GLM-5.3 Flash Uncensored: 100 replies, each 300 tokens long, with 4,000 tokens of context, come to $0.0532 at the 2026-11-16 catalog. The job table above names the first roleplay, security and abliterated rows by the same rule.
Roleplay from a phone
On a phone, roleplay goes through a web client: Janitor AI, with its proxy set to this base URL, or SillyTavern opened from the phone. The no-install route is /chat with a character or the Telegram bot, three free replies a day. A roleplay tune suits a card client because it replies straight away, with no reasoning pass first; the DeepSeek and GLM-5.3 builds reason before they answer and want a reply cap of at least 1,000 tokens.
| # | Roleplay tune | Out $/M | Window | 100 replies, 4K context |
|---|---|---|---|---|
| 1 | MythoMax L2 13B | $0.404 | 4K | $0.1737 |
| 2 | Manta Flash 1.0 | $0.56 | 16K | $0.0448 |
| 3 | Manta Mini 1.0 | $0.56 | 8K | $0.0448 |
Proxy URL, key and model id go in once, following the SillyTavern guide or the Janitor AI guide; KoboldCpp users have the KoboldCpp remote API guide. A phone and a desktop send an identical request, so the row above is the bill per reply either way.
Small uncensored LLMs, 27B and under
Lists written for local use rank small uncensored models by what a graphics card can hold: 8B, 12B, 24B and 27B at Q4. Over an API the size still matters, but for price and speed rather than VRAM, which is why the small builds sit at the bottom of the price column. Listed here: every build of 27B or less that passed the Nov 16 check. None of them carries the refusal layer a stock model ships with.
| # | Model | Build | Out $/M | Verdict |
|---|---|---|---|---|
| 1 | MythoMax L2 13B | Roleplay | $0.404 | Steady since Oct 20 |
| 2 | Cydonia 24B v4.3 | Roleplay | $0.60 | Steady since Oct 23 |
| 3 | Gemma 4 12B Starry Ocean | Roleplay | $0.875 | Steady since Oct 23 |
| 4 | Gemma 4 12B StationKeeper | Roleplay | $0.875 | Steady since Oct 23 |
| 5 | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | Abliterated | $1.0389 | Steady since Oct 23 |
Several of these have a UGI row under the same name, W/10 included; see the UGI table. If you would rather run the weights on your own card, the family pages link to the Hugging Face card of each build.
The 2026 picture
In 2026 the catalog holds 170 builds of four kinds: 38 abliterated, 37 uncensored, 78 roleplay, 7 security. 147 of them passed the Nov 16 check. The GLM-5.3, DeepSeek V4 and Kimi K3 edits are the youngest families, and they pushed the Llama and Mistral tunes of 2025 out of the first rows. The October 2026 list gives every build with the day it went on sale.
The rows move whenever the catalog moves, so what you are reading is the 2026 ranking as of the date printed on it. The month in the heading changes on the first of the month; the URL stays put.
UGI next to this board
DontPlanToEnd’s UGI Leaderboard on Hugging Face scores how uncensored a model is. UGI blends knowledge of sensitive subjects with W/10, willingness out of ten: how much pressure a model takes before it either refuses or wanders off the instructions. The questions are not published, and the board covers both open-weight and proprietary models.
What UGI leaves out is whether you can call the build today and what a token costs. That is what this page adds. 9 models on the catalog have a UGI row with the same author and model name. Highest UGI at the top:
| Model | UGI row | UGI score | W/10 | Verdict today | Out $/M |
|---|---|---|---|---|---|
| Anubis 70B v1.1 | TheDrummer/Anubis-70B-v1.1 | 50.79 | 8.5 | Changed Nov 15 | $1.085 |
| Cydonia 24B v4.3 | TheDrummer/Cydonia-24B-v4.3 | 44.52 | 7.0 | Steady since Oct 23 | $0.60 |
| Dolphin Mistral 24B Venice Edition | dphn/Dolphin-Mistral-24B-Venice-Edition | 40.79 | 7.8 | Steady since Oct 23 | $2.25 |
| Wayfarer Large 70B | LatitudeGames/Wayfarer-Large-70B-Llama-3.3 | 37.06 | 5.0 | Changed Nov 14 | $2.4535 |
| UnslopNemo 12B v4.1 | TheDrummer/UnslopNemo-12B-v4.1 | 34.17 | 6.2 | Steady since Nov 14 | $1.972 |
| Skyfall 36B v2 | TheDrummer/Skyfall-36B-v2 | 33.90 | 6.0 | Steady since Oct 23 | $2.80 |
| NeuralDaredevil 8B Abliterated | mlabonne/NeuralDaredevil-8B-abliterated | 33.20 | 6.5 | Changed Nov 14 | $1.76 |
| Mag Mell R1 12B | inflatebot/MN-12B-Mag-Mell-R1 | 31.14 | 6.0 | Changed Nov 15 | $1.7255 |
| Stheno 8B v3.2 | Sao10K/L3-8B-Stheno-v3.2 | 20.56 | 3.8 | Changed Nov 14 | $0.7035 |
How the ranking works
Two public inputs. No composite score.
First input: the real-model check
Every day the gateway sends the same fixed prompt to each model on the catalog and keeps the reply. The reply earns one of three verdicts. Steady means the reply matches how that build has always replied, and the date shows since when. Changed means the reply moved on the date given. Not checked yet means the model is new or missed the run.
What the check proves: the model behind the id is the one the id names, and it was answering when the run happened. It says nothing about how uncensored the model is: a stock model passes, and so does an abliterated edit, as long as each answers as itself. For that question, use UGI.
Second input: the live catalog
The catalog holds each model’s id, its build tags, its input and output price per million tokens and its context window. What the table shows is what the key is charged. A blank window means the catalog has no number for it.
The rule
- A Steady verdict today puts a model on the board. Rows marked Changed or Not checked yet go in the list under it.
- Within the board, a lower output price per million tokens ranks above a higher one. Output is the side that makes up most of a chat or roleplay bill.
- Equal output prices fall back to alphabetical order. Input price, window and build kind do not break ties.
Why price and not a score: price is the one figure in the row that is simply true. A benchmark depends on which prompts, which sampler and which judge. The price is the amount charged to the key, to the fraction of a cent, and it only moves when the catalog does.
Nobody can buy a row. No author-reported numbers, no sponsored placement. A model climbs by passing and costing less, and drops by failing or costing more.
Neither data file needs a key. Two commands rebuild the board from them:
# 1. Models that answered today's check as the build they are named forcurl -s https://unrestrictable.ai/proof.json \ | jq -r '.data[] | select(.verdict == "stable") | .model' > stable.txt # 2. Their output prices and context windows, lowest output price firstcurl -s https://unrestrictable.ai/pricing-index.json \ | jq -r --rawfile ok stable.txt '($ok | split("\n")) as $s | .data[] | select(.id | IN($s[])) | [.id, .output, .context_window] | @tsv' \ | sort -t$'\t' -k2,2nEvery model that passed
147 models answered the Nov 16 check as the build their id names. Output price sorts the table; any header re-sorts it. Each name opens the model’s page, where its first request is written out. Prices are per million tokens, matching the price list.
Off the board today
23 models are not ranked: 19 answered differently from before, and 4 have no verdict from this run. All of them still answer requests. No verdict is not a failed verdict.
| Model | Build | Out $/M | Verdict |
|---|---|---|---|
| Magidonia 24B v4.3 | Roleplay | $0.4235 | Changed Nov 14 |
| Cydonia 24B v2 | Roleplay | $0.4828 | Changed Nov 14 |
| Stheno 8B v3.2 | Roleplay | $0.7035 | Changed Nov 14 |
| Anubis 70B v1.1 | Roleplay | $1.085 | Changed Nov 15 |
| Anubis 70B v1 | Roleplay | $1.24 | Changed Nov 12 |
| Euryale 70B v2.2 | Roleplay | $1.428 | Changed Nov 14 |
| Cu-Mai R1 70B | Roleplay | $1.7255 | Changed Nov 14 |
| Hanami 70B x1 | Roleplay | $1.7255 | Changed Nov 15 |
| Mag Mell R1 12B | Roleplay | $1.7255 | Changed Nov 15 |
| Starcannon Unleashed 12B v1.0 | Roleplay | $1.7255 | Changed Nov 14 |
| NeuralDaredevil 8B Abliterated | Abliterated | $1.76 | Changed Nov 14 |
| Euryale 70B v2.3 | Roleplay | $1.972 | Changed Nov 15 |
| The Omega Directive 24B v2.0 | Roleplay | $2.00 | Changed Nov 14 |
| Wayfarer Large 70B | Roleplay | $2.4535 | Changed Nov 14 |
| Llama 3 70B Abliterated v3.5 | Abliterated | $2.80 | Changed Nov 12 |
| Qwen3.8 Max Uncensored | Uncensored | $3.00 | Changed Nov 15 |
| Kimi K3.1 Pro Uncensored | Uncensored | $5.82 | Changed Nov 14 |
| OrcaCyber Zero 1.5 | Security | $18.75 | Changed Nov 14 |
| Aion 3.5 Mini Uncensored | Uncensored | per request | Changed Nov 11 |
| Gemma 4 31B Mero Artemis v0.3.1 | Roleplay | $1.12 | Not checked yet |
| Rocinante 12B v1.1 | Roleplay | $2.38 | Not checked yet |
| Kimi K2.6 Uncensored | Uncensored | $7.34 | Not checked yet |
| Cyber Frontier | Security | $15.75 | Not checked yet |
The same rows, minus the verdict column, are on the models page.
What a long chat costs
Every turn of a chat sends the character card and the whole history again. The reply stays the same length; the input grows. This table prices 100 replies, 300 tokens each, at three context sizes for the first 12 priced rows that passed today. Same catalog read as everything else here.
| Model | 4,000 in | 8,000 in | 16,000 in |
|---|---|---|---|
| GLM-5.3 Flash Uncensored | $0.0532 | $0.0953 | $0.1795 |
| MythoMax L2 13B | $0.1737 | $0.3353 | $0.6585 |
| Manta Flash 1.0 | $0.0448 | $0.0728 | $0.1288 |
| Manta Mini 1.0 | $0.0448 | $0.0728 | $0.1288 |
| Cydonia 24B v4.3 | $0.2116 | $0.4052 | $0.7924 |
| GLM-5.3 Uncensored | $0.1382 | $0.251 | $0.4766 |
| Gemma 4 12B Starry Ocean | $0.0963 | $0.1663 | $0.3063 |
| Gemma 4 12B StationKeeper | $0.0963 | $0.1663 | $0.3063 |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | $0.4467 | $0.8623 | $1.6934 |
| Gemma 4 31B Split-Untied | $0.1715 | $0.3115 | $0.5915 |
| Veiled Calla 12B | $0.4515 | $0.8715 | $1.7115 |
| Gemma 4 31B Agares v1 | $0.4605 | $0.8889 | $1.7457 |
With 16,000 tokens in, 100 replies cost $0.1795 on GLM-5.3 Flash Uncensored and $1.7457 on Gemma 4 31B Agares v1. Every message the history adds widens that gap in dollars. That is why each model page suggests a Context Size for SillyTavern instead of the catalog maximum.
API or your own GPU
Local rankings sort uncensored models by VRAM: at Q4, roughly 5 GB for an 8B, 16 GB for a 27B, and 40 GB or more for a 70B. Run the weights yourself and there is no per-token bill and no prompt ever leaves the box; the cost is the hardware, plus the hours spent downloading, updating and swapping models.
| Run it locally | Call it here | |
|---|---|---|
| GPU | Yours, 20 GB or more for a 20B at Q4 | Not needed |
| Setup | LM Studio, Ollama or KoboldCpp, then a client | Open /chat, or put base URL, key and id in a client |
| Price | The card and the electricity | From $0.3685 per million output tokens |
| Models | Whatever fits in VRAM | 170 builds, one id apart |
| Proof it is the named model | The file on your disk | A dated verdict per model, every day |
Go local when prompts must never leave your hardware, or when a GPU is already sitting idle. Go API when there is no GPU, when you want to try several models by changing one string, or when one model has to serve a server, a phone and SillyTavern at the same time. The family pages record a Hugging Face card for the builds that have one, so the download is a click away from the API page.
The data files, and how to cite them
- pricing-index.json, or the same as pricing-index.csv: one row per model with id, name, build kind, context window, date, and the input and output price per million tokens.
- proof.json and proof.csv: each model’s verdict, the time it was checked, the steady-since date and the last change, as Unix milliseconds.
Both files refresh every five minutes and permit cross-origin reads, so a spreadsheet or a script can pull them straight in. When you cite the board, give the month and the check time:
Send your first request
Quickest: open /chat and pick the model. By API, the id in the model column goes in the request body. The format is the standard chat completions shape, so an OpenAI-compatible client needs only a new base URL and key.
curl https://api.unrestrictable.ai/v1/chat/completions \ -H "Authorization: Bearer sk-unr-live-…" \ -H "Content-Type: application/json" \ -d '{"model": "glm-5.3-flash-uncensored", "messages": [{"role": "user", "content": "Hello"}]}'The base URL is https://api.unrestrictable.ai/v1. The docs page has the settings for SillyTavern, Janitor AI, OpenCode and Cline.
One worked bill, same catalog read: 1,000 turns of 2,000 tokens in and 300 out, the shape of a long agent run, or of a fiction session with a full character card in context. On GLM-5.3 Flash Uncensored that is $0.32; on Kimi K3 Uncensored, the dearest row on today’s board, $20.09. Credit is prepaid and each request is settled at these rates.
Before a long job, look at the board again. Verdicts change once a day; prices change when the catalog does.
Questions
Two do. DontPlanToEnd's UGI Leaderboard on Hugging Face scores open-weight and proprietary models on what they know about sensitive topics and how willing they are to answer. This page ranks the models you can call by API: a daily real-model check first, output price second, with UGI's score shown for the 9 models UGI has tested under the same name. See the daily check.
Uncensored General Intelligence. The UGI Leaderboard is a Hugging Face Space that rates language models on knowledge of sensitive topics (hazardous, entertainment and socio-political) blended with W/10, a willingness score. It also publishes NatInt, a general knowledge and reasoning figure. The test questions stay private.
Open the Space and sort on UGI or W/10. Filters by model type (foundation, fine-tuned, merged, proprietary) narrow it down. To run a model from UGI without downloading it, look for its name on the models page here, then open it in /chat or put its id in a chat completions call. See the models page.
Willingness out of ten. It measures how far a model can be pushed before it refuses or stops following the instructions. W/10 is one component of UGI and has two halves: W/10-Direct for flat refusals and W/10-Adherence for drifting off the instructions.
GLM-5.3 Flash Uncensored is first on the November 2026 board: of the models that passed the Nov 16 check, it has the lowest output price, $0.3685 per million tokens. Among the rows that have a UGI score, Cydonia 24B v4.3 scores highest (UGI 44.52). The job table on this page names one model for each job.
GLM-5.3 Flash Uncensored is first on the November 2026 board: of the models that passed the Nov 16 check, it has the lowest output price, $0.3685 per million tokens. Among the rows that have a UGI score, Cydonia 24B v4.3 scores highest (UGI 44.52). The job table on this page names one model for each job. The board has no quality score of its own. It tells you which build answered as itself today, what a million output tokens cost, and UGI's figure where UGI has one. See the models by family.
On price, the first row today: GLM-5.3 Flash Uncensored at $0.3685 per million output tokens. It answers in /chat with nothing installed, or by API from a phone, a browser or a server. The job table names the first roleplay, security and abliterated rows the same way. See the browser chat.
MythoMax L2 13B: the roleplay tune that passed the Nov 16 check at the lowest price, $0.404 per million output tokens, so $0.1737 for 100 replies on 4,000 tokens of context. On the phone, run it in Janitor AI or SillyTavern with the proxy URL, key and model id entered once, or use /chat with no setup at all. See the Janitor AI guide.
MythoMax L2 13B: the build of 27B or under that passed the Nov 16 check at the lowest price, $0.404 per million output tokens. Small abliterated and fiction builds answer with no stock refusal layer, and over an API their size sets the price rather than the hardware you need.
From today's board, MythoMax L2 13B: the roleplay tune that passed the Nov 16 check at the lowest price, $0.404 per million output tokens. A roleplay tune replies at once with no reasoning pass, and that suits a card client. The DeepSeek and GLM-5.3 builds reason first and need a reply cap of at least 1,000 tokens. See the SillyTavern guide.
GLM-5.3 Flash Uncensored: $0.3685 per million output tokens on the 2026-11-16 catalog, which makes 100 replies of 300 tokens on 4,000 tokens of context $0.0532. Input is charged on top at the build's own rate.
Heretic is an open-source tool that does abliteration automatically: it finds the refusal direction in a model and strips it from the weights. A Heretic build is therefore an abliterated build produced by that tool. 9 builds on the catalog have Heretic in the id, 9 of them on today's board. The catalog tags them abliterated. See abliterated vs uncensored.
It is. Dolphin is Cognitive Computations' family of uncensored fine-tunes. Dolphin Mistral 24B Venice Edition is on the catalog as dolphin-mistral-24b-venice, $2.25 per million output tokens, and today's verdict is steady since oct 23. UGI has a row for it under the same name.
No. Open /chat and the model answers in the browser. By API, every model on the board sits behind the base URL https://api.unrestrictable.ai/v1: paste a key and send the id from a phone, a server or a roleplay client. Credit is prepaid from $10 and charged per token, with nothing to download. See the client setup docs.
Yes. Artificial Analysis runs an Image Arena leaderboard as a Hugging Face Space (read October 25, 2026). The uncensored image and video models you can call here are on the media page with per-image and per-second prices. See the media page.
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.