Uncensored AI for coding

Short answer: Obscurify.ai serves uncensored and decensored models with tool calling and up to 1M tokens of context, through an OpenAI-compatible API (https://obscurify.ai/v1) and an Anthropic-compatible one for Claude Code. Point your coding agent at it and change the model id: we tested the setup in Claude Code, Cline, OpenCode, Aider, Kilo Code and Continue on November 8, 2026. The cheapest uncensored coding model, GLM 5.3 Flash Uncensored, costs $0.22 per million input tokens. Uncensored models need Pro ($45/month with $35 of usage included); the decensored and cybersecurity models work on any account. The account is a random 16-digit number, with no email. New accounts get $1 of starter credit, with the Pro models (uncensored ones included) for their first 7 days, while this month’s promo budget lasts.

Why use an uncensored model for code

Mainstream coding assistants refuse more than you’d expect, and an agent that stops halfway through a task to lecture you is worse than one that never started. The refusals cluster around a few kinds of work:

• Security code on your own systems. Exploit proofs of concept for a CVE you’re patching, fuzzers, scanners, a reverse shell for a CTF box, a SQL injection test against your own login. In our refusal test, Claude Sonnet 5.5 refused both of the security prompts every uncensored model answered.
• Reverse engineering. Reading a binary you own, cracking your own license check, patching a game save. Claude refusing the popular Reverse Engineer Anything tool is how a lot of people end up here.
• Scrapers and automation that a model decides look like abuse.
• Code that sits next to adult or violent content, from a game’s dialogue system to a moderation classifier that needs examples.

An uncensored model does the task and leaves the judgment to you. The terms still apply: no writing or spreading malware, and no breaking into systems you aren’t authorized to access.

Which models to use

A coding agent needs tool calling (to read files, edit them and run commands) and a long context (agents send whole files and long histories). These uncensored models have both, read live from the catalogue:

• Abliterated Large Flash $1.10 / $2.21 per 1M · 1M context · Pro plan
• Abliterated Model Large $3.31 / $5.53 per 1M · 1M context · Pro plan
• Abliterated Model Large V2 $3.31 / $5.53 per 1M · 1M context · Pro plan
• Llama 3.1 8b (uncensored) $0.88 / $1.77 per 1M · 32K context · Pro plan
• DeepSeek V4.1 Flash 0910 Uncensored $0.22 / $0.88 per 1M · 512K context · Pro plan
• DeepSeek V4.1 Flash 0910 Uncensored Thinking $0.22 / $0.88 per 1M · 512K context · Pro plan
• DeepSeek V4.1 Flash 1017 Uncensored $0.22 / $0.88 per 1M · 512K context · Pro plan
• DeepSeek V4.1 Flash 1017 Uncensored Thinking $0.22 / $0.88 per 1M · 512K context · Pro plan
• DeepSeek V4.1 Flash Uncensored $0.22 / $0.88 per 1M · 512K context · Pro plan
• DeepSeek V4.1 Flash Uncensored Thinking $0.22 / $0.88 per 1M · 512K context · Pro plan
• Gemma 4 26B A4B Chimera X Heretic $0.13 / $0.42 per 1M · 262K context · Pro plan
• Gemma 4 26B A4B Uncensored TEE $0.19 / $0.91 per 1M · 262K context · Pro plan
• CyberHeabsy $1.66 / $4.97 per 1M · 1M context · Pro plan
• Qwen 3.8 27B Uncensored Fast by Heabsy $1.10 / $3.31 per 1M · 1M context · Pro plan
• DeepSeek V4.1 Flash Heretic $0.22 / $0.88 per 1M · 512K context · Pro plan
• DeepSeek V4.1 Flash Heretic Thinking $0.22 / $0.88 per 1M · 512K context · Pro plan
• Qwen 3.8 27B Uncensored $1.10 / $3.31 per 1M · 1M context · Pro plan
• Qwen 3.8 27B Uncensored Thinking $1.10 / $3.31 per 1M · 1M context · Pro plan
• Qwen3.8 27B Uncensored TEE $0.39 / $1.95 per 1M · 262K context · Pro plan
• Qwen3.8 27B Uncensored TEE (Thinking) $0.39 / $1.95 per 1M · 262K context · Pro plan

And 5 more on the model list.

Our picks, from the refusal test and from running them in agents:

Be honest with yourself about the tradeoff. Removing refusals doesn’t make a model a better programmer, and the frontier models (Claude, GPT, Gemini, also on Obscurify.ai) still write better code on hard refactors. A practical setup is a frontier model for everyday work and an uncensored one for the tasks it refuses. Both live behind the same key.

Set it up in your agent

Each of these was set up against the live API. “Tested” means we ran the agent and it edited a file through a tool; “docs-checked” means we verified the current config format against the tool’s own docs.

Anything else that takes an OpenAI base URL works the same way: base URL https://obscurify.ai/v1, your API key, and the full model id including the maker prefix (z-ai/..., not glm-5.3-flash-uncensored).

Security and reverse-engineering tools

The same models drive security tooling, for authorized work: pentests in scope, CTFs, bug bounties, your own binaries, malware analysis.

• PentestGPT on a cybersecurity model
• Kali GPT: shell-gpt in your Kali terminal
• Ghidra MCP and IDA Pro MCP driven by a coding agent
• Reverse Engineer Anything without the refusals

In our test, OpenCode on the decensored DeepSeek model disassembled a crackme we compiled, pulled the password out of its read-only data and ran the binary to confirm it. The cybersecurity LLM page compares the security-tuned models.

What happens to your code

Agents send a lot: whole files, terminal output, sometimes secrets you forgot were in a config file. So it’s worth knowing where it goes.

• Prompts and replies, your code included, aren’t stored on Obscurify.ai’s servers.
• Your identity is scrambled before a request reaches the model provider. The provider sees the code (it needs it to answer), not who sent it.
• Models with the ZDR tag only go to zero-data-retention routes, whose providers’ published policies say they don’t train on or keep the data. These rest on published policies, not signed agreements. The uncensored and Heretic models in the table above have the tag; the OrcaCyber model doesn’t.
• The account is a random 16-digit number, and you can pay with Monero or Bitcoin, so no request leads back to a name.

If code must never leave your machine at all, run a local model. For everything else, this keeps it private without running the models yourself.

Uncensored coding AI questions

What is the best uncensored model for coding?

For agent work, one with tool calling and a long context. GLM 5.3 Flash Uncensored is the cheapest that answered every prompt in our refusal test, and DeepSeek V4.1 Flash Uncensored is the one we ran end to end in Claude Code. On any account without Pro, the decensored DeepSeek V4.1 Flash 1017 Heretic had the cleanest tool calls.

Can I use an uncensored model in Claude Code?

Yes. Obscurify.ai serves the Anthropic Messages API, so set ANTHROPIC_BASE_URL=https://obscurify.ai, your API key and ANTHROPIC_MODEL to the model id. The Claude Code tutorial has the full setup.

Do I need the Pro plan?

For the uncensored models, yes. The decensored and cybersecurity models, which also skip most refusals, work on every plan, including Pay As You Go with no monthly fee.

Is my code used for training?

Not by us. Models with the ZDR tag only go to zero-data-retention routes whose providers’ published policies say they don’t train on or keep the data; other models follow their provider’s own policy.

Will it write anything I ask?

No. It answers far more than mainstream assistants, but the terms still apply: no malware, and no attacking systems you aren’t authorized to test.

Last updated 2026-11-08. Plan prices and model lists on this page are read live from the site.