---
title: Uncensored vs Abliterated AI Models: Differences & API IDs
description: Compare uncensored and abliterated models, prompt-level and weight-level changes, exact DeepSeek, GLM and Qwen IDs, quality and cost.
canonical: https://routerhub.si/guides/uncensored-vs-abliterated
---

Model guide
# Uncensored vs Abliterated AI Models: Differences & API IDs
Uncensored models use reduced-refusal prompting or training; abliterated models use weight-level refusal reduction.
By RouterHub · Updated 3 November 2026

## What is the difference?
Uncensored models are reduced-refusal language model variants for open-ended generation, coding and creative work. Uncensored variants use prompting, fine-tuning, weight changes, or a combination of these techniques.
Abliteration reduces refusal-related behavior by modifying model weights. RouterHub serves uncensored and abliterated variants under separate API IDs, with their own token prices and context limits.
Uncensored and abliterated models compared| Question | Uncensored | Abliterated |
| What is it? | A reduced-refusal model variant | A model with weight-level refusal reduction |
| How is it configured? | Prompt profiles, fine-tuning or weight changes | Changes to refusal-related weight representations |
| What should you compare? | Exact ID, instruction handling, quality and cost | Exact ID, instruction handling, quality and cost |

## Which should you choose?
Choose an uncensored model for a reduced-refusal instruction profile. Choose an abliterated model for weight-level refusal reduction. Compare each variant on your prompts, required output format and cost per successful result.
RouterHub’s prompt-profile uncensored variants use a server-managed instruction profile. Customer system/developer instructions can be replaced by that profile; put task-specific requirements in the user message.

## Uncensored and abliterated API IDs

  RouterHub token prices, USD per 1 million tokens
  | Model / exact API ID | Input / 1M | Cached / 1M | Output / 1M | Context | Actions |

  | DeepSeek V4 Pro Uncensored Uncensored deepseek-v4-pro-uncensored | $0.20 −00% | $0.0016 | $0.40 | 1M | Chat |
| DeepSeek V4.1 Flash Uncensored Uncensored deepseek-v4.1-flash-uncensored | $0.30 $0.15 −50% | $0.0027 | $1.20 $0.60 | 1M | Chat |
| GLM 5.3 Flash Uncensored Uncensored glm-5.3-flash-uncensored | $0.15 −00% | $0.013 | $0.50 | 1M | Chat |
| GLM 5.3 Uncensored Uncensored glm-5.3-uncensored | $1.40 $0.20 −85% | $0.12 | $4.40 $0.63 | 1M | Chat |
| GLM 5.3 Flash Abliterated Abliterated glm-5.3-flash-abliterated | $0.30 −00% | $0.12 | $1.00 | 1M | Chat |
| GLM 5.3 Abliterated Abliterated glm-5.3-abliterated | $1.40 $0.40 −70% | $0.16 | $4.40 $1.33 | 1M | Chat |
| Qwen 3.8 Max Uncensored Uncensored qwen3.8-max-uncensored | $2.00 $0.25 −87% | $0.031 | $6.00 $0.75 | 1M | Chat |
| Kimi K3.1 Pro Uncensored Uncensored kimi-k3.1-pro-uncensored | $2.00 $0.30 −85% | $0.030 | $10.00 $1.50 | 1M | Chat |
| DeepSeek V4.1 Flash Abliterated Abliterated deepseek-v4.1-flash-abliterated | $0.30 −00% | $0.055 | $1.25 | 1M | Chat |
| DeepSeek V4 Pro Abliterated Abliterated deepseek-v4-pro-abliterated | $0.40 −00% | $0.040 | $1.60 | 1M | Chat |
| DeepSeek V4.1 Flash 1017 Abliterated Abliterated deepseek-v4.1-flash-1017-abliterated | $0.30 −00% | $0.055 | $1.25 | 1M | Chat |

## Compare cost per successful task
For ordinary uncached text usage, estimate cost as `input tokens × input price / 1,000,000 + output tokens × output price / 1,000,000`. Reported cache reads and writes use their own published rates. The final Usage record, not time spent waiting or a headline token limit, is the billing reference.
Example: at $0.20 per million input tokens and $0.60 per million output tokens, 10,000 uncached input tokens and 2,000 output tokens cost an estimated $0.0032. This is an illustrative calculation, not a quoted price for a particular model.
For model selection, divide the total cost of all measured attempts by the number of acceptable results. A lower token rate is not automatically a lower cost per completed task.

## Test refusal behavior and quality separately

- Fix the task. Use the same inputs, output format, reasoning setting and output budget for each exact model ID.
- Separate refusal from failure. Record an explicit refusal separately from a timeout, empty response, invalid JSON, wrong answer or truncated generation.
- Keep every attempt. Repeat your tests and retain unsuccessful outputs, not just the best answer.
- Measure useful work. Compare task acceptance, instruction following, tool-call validity and cost per successful task, not token price alone.
- Re-test changes. A hosted alias or model update can change behavior even when your client configuration stays the same.
RouterHub’s [public coding evaluations on GitHub](https://github.com/routerhubSI/routerhub-model-evals) include prompts, raw outputs, settings, candidate hashes and sandbox-scored tests. The GLM results record both the default-effort attempts and the low-effort cohort, so you can compare runs with their original settings.

## Next steps

- [Use an uncensored model through an API Connect with Python or cURL.](https://routerhub.si/guides/uncensored-ai-api)
- [How abliteration works Understand weight-level refusal reduction.](https://routerhub.si/guides/what-is-abliteration)
- [RouterHub comparison repository Read the source, examples and project notes on GitHub.](https://github.com/routerhubSI/uncensored-vs-abliterated-ai)