Abliterated AI models

Models with refusal behaviour reduced directly in their weights. Each one has its own API ID and price, on the same key and balance as every other model.

Models
5
Input from
$0.30/1M
Longest context
1M
Families
DeepSeek, Z.ai

Compare abliterated models

Cheapest first. Crossed-out prices are the official rate for the standard base model.

Abliterated AI models on RouterHub, USD per 1M tokens
Model / exact API IDInput / 1MCached / 1MOutput / 1MContextActions
DeepSeek V4.1 Flash Abliterateddeepseek-v4.1-flash-abliterated $0.30 $0.055 $1.25 1M ChatConnect
DeepSeek V4.1 Flash 1017 Abliterateddeepseek-v4.1-flash-1017-abliterated $0.30 $0.055 $1.25 1M ChatConnect
GLM 5.3 Flash Abliteratedglm-5.3-flash-abliterated $0.30 $0.12 $1.00 1M ChatConnect
DeepSeek V4 Pro Abliterateddeepseek-v4-pro-abliterated $0.40 $0.040 $1.60 1M ChatConnect
GLM 5.3 Abliteratedglm-5.3-abliterated $1.40$0.40−70% $0.16 $4.40$1.33 1M ChatConnect

How they work

What abliteration is

A weight-level edit that removes the direction a model uses to refuse. The model's instructions are left as they are.

How it differs

Uncensored variants change the instructions or training; abliterated ones change the weights. Quality can shift either way, so test on your own task.

How they're billed

Per token from your prepaid balance, like every other model. The same API key works for these and every standard model.

Choose one for your task

Run the same prompts across two or three models from the table, then compare.

  1. Fix the taskUse the same inputs, output format, reasoning setting and output budget for every model ID you compare.
  2. Separate refusal from failureCount an explicit refusal apart from a timeout, empty reply, invalid JSON or wrong answer.
  3. Keep every attemptRepeat each test and keep the failed outputs, not only the best answer.
  4. Measure useful workCompare cost per accepted result, not token price alone.

Guides

By family: DeepSeekGLMQwenKimi