A sourced price comparison of hosted APIs that serve abliterated or uncensored open-weight models, plus copy-paste configs for Claude Code, Codex, OpenCode and the OpenAI SDK.
Prices change without notice. Every number below was read on 2026-10-27 from the source linked in the Sources section. If you find a stale price, open an issue or a PR with the new source.
Disclosure. The Refuseless links in this repo are affiliate links: if you sign up through them and deposit at least $10, you get a $5 bonus and I earn a commission. Other providers are listed because they are relevant, and their links carry no referral code. Nobody paid for placement.
Abliteration finds the direction in the model's activations that drives refusals and removes it from the weights. It differs from a jailbreak, which injects text into the prompt and has to be reapplied every session.
It adds no knowledge. An abliterated model is as capable as its base model at best, and a sloppy edit makes it worse. Judge a build by its base model's benchmarks and by testing it on your own tasks.
"Uncensored" in a provider's catalog does not always mean abliteration: some providers do not say whether the refusals were removed from the weights, by fine-tuning or by a permissive system prompt. Ask before you pay.
US$ per 1M tokens, input / output.
| Provider | Model id | Input / output | Context | API formats | How refusals were removed |
|---|---|---|---|---|---|
| Refuseless | glm-5.3-abliterated |
1.25 / 2.25 | 1,048,576 | OpenAI, Anthropic | Weight edit (abliteration), per the provider |
| NanoGPT | z-ai/glm-5.3-uncensored |
1.25 / 2.25 (cache read 0.30) | 1,048,576 | OpenAI | Not stated ("experimental uncensored") |
| abliteration.ai | abliterated-model-large-v2 |
3.00 / 5.00 | ~1.05M | OpenAI, Anthropic | Abliteration |
| Venice | Abliterated Large V2 | 3.00 / 5.00 | ~1.05M | OpenAI | Not stated |
| Provider | Model | Input / output | Notes |
|---|---|---|---|
| NanoGPT | GLM 5.3 Flash Uncensored | 0.20 / 0.80 | Output limit 32,768 |
| NanoGPT | GLM 5.3 Flash Cybersecurity | 0.15 / 0.50 | Uncensored base with moderation for illegal activity |
| MagmaRouter | glm-5.3-flash-xploded |
0.30 / 1.00 | OpenAI and Anthropic formats, per-key spend cap |
| Venice | DeepSeek V4.1 Flash Uncensored | 0.33 / 1.32 | |
| Venice | Abliterated Large Flash | 1.25 / 2.50 | |
| Refuseless | glm-5.3-flash-abliterated |
Not published | Price not listed in the public docs |
| OrcaRouter | OrcaCyber GLM 5.3 Flash | Not published per token | Publishes its weights and refusal-rate method on Hugging Face |
For price alone on the Flash tier, NanoGPT and MagmaRouter are the cheapest options listed here.
OpenRouter listed no abliterated or uncensored model in its public /api/v1/models
catalog on 2026-10-27.
Each file in configs/ points one tool at Refuseless. For another provider, swap the base URL,
the key and the model id.
| Tool | File | Wire format |
|---|---|---|
| Claude Code | configs/claude-code.md |
Anthropic Messages |
| Codex CLI | configs/codex.toml |
OpenAI Responses |
| OpenCode | configs/opencode.json |
OpenAI Chat Completions |
| OpenAI SDK (Python) | configs/openai_sdk.py |
OpenAI Chat Completions |
Removing refusals moves the judgment to you. Every provider above forbids attacking systems you do not own or are not authorized to test, and so does the law in most places.
Read on 2026-10-27.
- Refuseless: docs.refuseless.com (models, SDKs, OpenAPI spec) and the pricing shown on refuseless.com.
- NanoGPT: nano-gpt.com/pricing and the GLM 5.3 Uncensored model page.
- abliteration.ai and MagmaRouter: magmarouter.com/alternatives/abliteration-ai.
- Venice: docs.venice.ai/overview/pricing.
- OrcaRouter: orcarouter.ai model pages and huggingface.co/orcarouter.
- OpenRouter:
GET https://openrouter.ai/api/v1/models.
MIT.