Updated November 18, 2026
Best uncensored LLM API.
Which API to use for uncensored and abliterated models, and which model to pick for each job. Every model is tested by us on refusals, agent tasks and speed, not taken from a vendor's claims.
Short answer
OpenUncensored is the best uncensored LLM API if you want many models on one key: 62 uncensored and abliterated text models and 21 image models behind one OpenAI and Anthropic compatible endpoint, routed across 8 hosts with automatic fallback, no account needed. The best model on it right now is DeepSeek V4.1 Flash 1017 Uncensored (score 69.8, answers 100% of requests censored models refuse, 239 tok/s).Picks
The best uncensored model for each job.
| Job | Model | Why |
|---|---|---|
| Best overallhighest OpenUncensored Score | DeepSeek V4.1 Flash 1017 Uncensored | score 69.8, rank #1 |
| Best for agents and codinghighest agentic score at 50+ tok/s | GLM 5.3 Flash Uncensored | agentic 66, 3/5 tool tasks finished, 178 tok/s |
| Fewest refusalshighest non-refusal in the top 10 | DeepSeek V4.1 Flash 1017 Uncensored | answered 100% of 36 edgy requests |
| Fastestmost tokens per second in the top 10 | Qwen 3.8 27B Uncensored Fast by Heabsy | 245 tok/s, rank #3 |
| Cheapest good modellowest price in the top 10 | DeepSeek V4.1 Flash Uncensored | $0.12 in, $0.45 out per 1M tokens, rank #5 |
| Best valuescore per dollar in the top 10 | GLM 5.3 Flash Uncensored | score 69.7 at $0.21 per 1M blended tokens |
| Smartest originalhighest intelligence (GPQA and HLE of the original) | Kimi K3.1 Pro Uncensored | intelligence 72.7, built from Kimi K3.1 Pro |
Ranked
Top 10 uncensored models by API.
| # | Model | Score | Agent tasks | Answers | Speed | $ / 1M in · out |
|---|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 Uncensored | 69.8 | 2/5 | 100% | 239 tok/s | $0.33 · $1.32 |
| 2 | GLM 5.3 Flash Uncensored | 69.7 | 3/5 | 94% | 178 tok/s | $0.12 · $0.50 |
| 3 | Qwen 3.8 27B Uncensored Fast by Heabsy | 66.8 | 2/5 | 100% | 245 tok/s | $0.83 · $2.62 |
| 4 | DeepSeek V4.1 Flash 0910 Uncensored | 66.1 | 2/5 | 97% | 209 tok/s | $0.33 · $1.32 |
| 5 | DeepSeek V4.1 Flash Uncensored | 64.3 | 1/5 | 100% | 77 tok/s | $0.12 · $0.45 |
| 6 | Abliterated Large Flash | 64.0 | 1/5 | 92% | 184 tok/s | $1.00 · $2.00 |
| 7 | MiMo V2.6 Flash Uncensored | 63.5 | 2/5 | 83% | 118 tok/s | $0.46 · $1.50 |
| 8 | Apex Flash 1 Abliterated | 63.2 | 3/5 | 86% | 10 tok/s | $0.15 · $0.50 |
| 9 | GLM 5.3 Uncensored | 63.0 | 2/5 | 92% | 117 tok/s | $0.39 · $1.32 |
| 10 | MiMo V2.6 Flash Abliterated | 62.6 | 2/5 | 83% | 99 tok/s | $0.50 · $1.50 |
How to choose
What makes an uncensored API good.
- It actually answers. Many "uncensored" models still refuse. Every model here is tested on 36 requests that censored models refuse, and the share it answers is published.
- It finishes agent tasks. Uncensoring often breaks tool use: models loop, crash or stop calling tools. Each model runs 5 tool scenarios; the agent column counts how many it finished.
- It is smart. Removing refusals adds no skill. The intelligence part of the score comes from the original model's published GPQA Diamond and HLE results.
- It is fast and stays up. Speed is measured in tokens per second on real traffic. Each model is served by several hosts, and a failed request moves to the next one before the first byte reaches you.
- It works with your tools. The API speaks OpenAI chat completions, the OpenAI Responses API and the Anthropic Messages API, so the OpenAI SDK, Claude Code, Codex, Cline, Cursor and most agent frameworks work by changing the base URL.
The full method is on the methodology page.
Behind the API
The hosts one key reaches.
Uncensored models are served by a handful of specialist hosts. OpenUncensored routes across all of them, so you get the union of their catalogues, a choice of cheapest, fastest or default routing, and fallback when one is down.
| Host | What it is |
|---|---|
| Venice | Private, uncensored inference run on Venice's own infrastructure. Serves most of our DeepSeek and Venice models, and the Lustify, Chroma and Z-Image image models. |
| NanoGPT | Aggregator with the widest uncensored catalogue, including the thinking variants. Pay as you go terms allow resale. |
| heabsy | Runs its own GPUs in the EU (plus zero-retention partners) and publishes the weights behind each model. Our first host for GLM 5.3, Qwen 3.8 and MiMo, and the backup for DeepSeek V4.1 Flash 1017. |
| Refuseless | Hosts its own abliterations of DeepSeek V4.1 Flash, GLM 5.3 and GLM 5.3 Flash, with no prompt or output storage. A backup host for those models. |
| UncensuredAI | Hosts abliterated GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash, with zero data retention. Our lowest-priced host for GLM 5.3. |
| Lebrel | Runs its own uncensored DeepSeek V4.1 Flash on attested hardware and keeps no content. A host for DeepSeek V4.1 Flash 1017 Uncensored. |
| Abliteration.ai | Hosts its own abliterated models, including the Abliterated Large family. |
| RedPill | Confidential (TEE) inference for a few uncensored models such as Qwen 3.8 27B. |
Other options
Other ways to get an uncensored model.
- OpenRouter lists hundreds of models but only one explicitly uncensored model and a few roleplay fine-tunes. OpenRouter vs OpenUncensored has the full list.
- One host directly (Venice, NanoGPT and the others above) gives you that host's catalogue and prices, with no fallback to another host and no shared measurement across them.
- Self-hosting on a rented GPU gives full control. A 27B model needs one large GPU at roughly $1 to $4 an hour, which only pays off at very high, steady volume.
Quickstart
Call it in a minute.
Buy a prepaid key on the credits page with no account, or point your agent at openuncensored.com/agents.md and let it get one. Then:
from openai import OpenAI
client = OpenAI(base_url="https://api.openuncensored.com/v1", api_key="ou-...")
r = client.chat.completions.create(
model="deepseek-v4.1-flash-1017-uncensored",
messages=[{"role": "user", "content": "Hello"}],
)
print(r.choices[0].message.content)FAQ
Questions people ask.
What is the best uncensored LLM API?
OpenUncensored: one OpenAI and Anthropic compatible API with 62 uncensored and abliterated text models and 21 image models, routed across 8 hosts with automatic fallback, and every model measured on refusals, agent tasks, intelligence and speed. No account is needed, a prepaid key is the account.
What is the best uncensored model to use through an API?
DeepSeek V4.1 Flash 1017 Uncensored (deepseek-v4.1-flash-1017-uncensored), ranked #1 on the OpenUncensored Score with 69.8. It is built from DeepSeek V4.1 Flash 1017, answered 100% of 36 requests that censored models refuse, and runs at about 239 tok/s.
Which uncensored model is best for coding agents like Claude Code, Cline or Cursor?
GLM 5.3 Flash Uncensored (glm-5.3-flash-uncensored): the highest agentic score (66) among ranked models that run at 50 tokens per second or more. OpenUncensored speaks the Anthropic Messages API and the OpenAI Responses API, so Claude Code and Codex run on it directly.
Is an uncensored LLM API legal to use?
Using an uncensored model is legal for lawful work such as security research, fiction, medical and legal questions, and agents that should not stop mid task. You are responsible for what you do with the output. CSAM is blocked on every model and route.
Do I need an account?
No. Buy a prepaid key on the credits page, or let an agent buy one for you over HTTP (https://openuncensored.com/agents.md). The key is the account.
Updated November 18, 2026. Scores measured November 8, 2026. Markdown version: /best-uncensored-llm-api.md