Abliterated models APIPaid credit only

Abliterated models API. Refusals removed, rules stated.

An abliterated model is an open-weight LLM with its refusal behaviour removed from the weights, so it answers some requests the original model declines. RouterPlex serves 5 of them through the same OpenAI- and Anthropic-compatible API as every other model, from $0.12 per 1M input tokens, at RouterPlex-set rates. Read what abliteration is for the background.

Prices read from the live catalog on 17 November 2026.

Abliterated routes17 November 2026

5

Routes

$0.12

Cheapest input / 1M

Request storage 1 to 365 days

Live catalog

Every abliterated route, priced.

USD per 1M tokens, read from the catalog on each request. Each rate is a RouterPlex-set rate, not the model maker's own price. Sorted by the cost of a 100K-input, 10K-output request.

Model idBase modelMethodContextMax outputInput / 1MCached input / 1MOutput / 1MSupports
GLM-4.7 Flash Abliterated

glm-4.7-flash-abliterated

GLM-4.7 FlashZhipuWeight edit200K24,000$0.12$0.09$0.60text only, tools, reasoning
Gemma 4 26B Abliterated

gemma-4-26b-abliterated

Gemma 4 26BGoogleFine-tune256K8,192$0.25not listed$0.75images, tools
GLM-5.3 Flash Abliterated

glm-5.3-flash-abliterated

GLM-5.3 FlashZhipuWeight edit1M131,072Was $0.50, now $0.25, 50% offWas $0.30, now $0.06, 80% offWas $1.60, now $1.00, 38% offimages, tools, reasoning
DeepSeek V4.1 Flash Abliterated

deepseek-v4.1-flash-abliterated

DeepSeek V4.1 FlashDeepSeekWeight edit1M131,072Was $0.70, now $0.50, 29% offWas $0.20, now $0.07, 65% off$1.50images, tools, reasoning
GLM-5.3 Abliterated

glm-5.3-abliterated

GLM-5.3ZhipuWeight edit1M131,072Was $4.00, now $2.00, 50% offWas $0.70, now $0.30, 57% offWas $6.00, now $4.00, 33% offimages, tools, reasoning
  • glm-5.3-flash-abliterated: price cut 30 October 2026, input 50% lower, output 38% lower, cached input 80% lower. RouterPlex lowered the rate after its upstream cost fell.
  • deepseek-v4.1-flash-abliterated: price cut 30 October 2026, input 29% lower, cached input 65% lower. RouterPlex lowered the rate after its upstream cost fell.
  • glm-5.3-abliterated: price cut 30 October 2026, input 50% lower, output 33% lower, cached input 57% lower. RouterPlex lowered the rate after its upstream cost fell.

"Not listed" means the catalog shows no separate cached-input rate for that route.

"Weight edit" routes had refusals edited out of the weights (abliteration). "Fine-tune" routes were trained further so they stop refusing. Both are served on RouterPlex through a third-party host.

Who uses them

Security and safety work stock models refuse.

Stock models often decline legitimate security questions. These routes answer them. The work must still be lawful and authorised, and the Terms below apply to every request.

  • Authorised security testing

    Penetration-test planning and reporting for systems you own or are authorised in writing to test. RouterPlex may ask for evidence of that authorisation.

  • Red teaming your own AI systems

    Running attack prompts against your own LLM apps and guardrails, and building evaluation sets to measure them.

  • Malware analysis education

    Reading and explaining suspicious scripts, logs and disclosed CVEs, then writing detection rules and patches. Creating malware or exploits for use against systems you are not authorised to test is prohibited.

  • Content-moderation research

    Classifying and labelling harmful text for your own moderation filters. The ban on sexual and NSFW content applies to these models in full.

The rules

Read these before the first request.

Quoted from the Terms of Service and the Privacy Policy. If this page and those documents ever differ, the Terms and the Privacy Policy apply. Sending a request to an abliterated model accepts the storage terms for that request.

Paid credit only

“Promotional credit, including top-up match bonuses, vouchers, sign-up and volume bonuses, cannot be used on abliterated models; those models draw only on credit you have paid for. Promotional credit is used on other models first.”

Terms, section 03

Requests are stored

“We store the requests you send to abliterated models, together with your account identifier, the model name, and the time of the request, for security, abuse prevention, and enforcement of these Terms. Stored requests are kept for at least 1 day and no more than 365 days, and are then deleted. Access is restricted to RouterPlex. Requests to all other models are not stored.”

Terms, section 05

Prohibited uses

“In addition to section 04, you must not use abliterated models to: design, develop, build, acquire, or use weapons, including explosives and chemical, biological, radiological, or nuclear weapons; test, scan, attack, or gain access to any system, network, application, or account that you do not own or are not explicitly authorised in writing to test, or create malware or exploits for use against such systems.”

Terms, section 05

Authorised security work

“Security research and penetration testing are permitted only against systems you own or are authorised to test, and we may ask you for evidence of that authorisation.”

Terms, section 05

No sexual or NSFW content

“The following rules apply to every model on the Service, including abliterated models, open-weight models, and models with reduced or no content filtering. You must not use the Service to: generate sexual or sexually explicit content of any kind, in text or images, including pornography, erotica, explicit sexual role-play, and sexual descriptions of real or fictional people; [...] build or operate NSFW, adult, AI companion, romantic, or relationship chatbots or characters.”

Terms, section 04

Bans and forfeited balance

“If you violate this section, we may suspend or permanently terminate your account without notice. Any remaining balance is forfeited and will not be refunded; all payments are final.”

Terms, section 05

No refunds

“All sales of credit are final: top-ups and plan payments are non-refundable, whether the credit is used or unused, except where a refund is required by applicable law or granted by our payment provider under its buyer terms.”

Terms, section 03

Privacy Policy

“This applies only to requests that name an abliterated model in the model field; no other model’s traffic is logged. If you do not want your requests stored, do not use the abliterated models.”

Privacy Policy, section 02

In practice: the $1 trial key gets HTTP 402 (paid_model_required) on these routes, so add a paid top-up first. A hard spend cap on the key stops a long run at your limit.

First request

Call deepseek-v4.1-flash-abliterated with curl or the OpenAI SDK.

The OpenAI base URL is https://api.routerplex.com/v1 and the Anthropic base URL is https://api.routerplex.com. One key covers both. Swap the model id for any id in the table; ids keep their dots, page URLs use hyphens.

curl /chat/completions
curl https://api.routerplex.com/v1/chat/completions \
-H "Authorization: Bearer $ROUTERPLEX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4.1-flash-abliterated",
"messages": [
{"role": "user", "content": "Explain how CVE-2021-44228 (Log4Shell) works and write a detection rule for JNDI lookup strings in web server logs."}
]
}'
OpenAI SDK (Python)
import os
from openai import OpenAI
 
client = OpenAI(
base_url="https://api.routerplex.com/v1",
api_key=os.environ["ROUTERPLEX_API_KEY"],
)
reply = client.chat.completions.create(
model="deepseek-v4.1-flash-abliterated",
messages=[{
"role": "user",
"content": "Explain how CVE-2021-44228 (Log4Shell) works and write a "
"detection rule for JNDI lookup strings in web server logs.",
}],
)
print(reply.choices[0].message.content)

How to choose

Pick by the job, not by a ranking.

Each rule reads only catalog fields: price, context, output limit, image input and cached-input rate. They are not quality rankings, so test your own prompts on two routes before a large run.

Lowest bill per request

glm-4.7-flash-abliterated has the lowest cost for a 100K-input, 10K-output request: $0.018. Use it for large prompt sets where volume drives the bill.

GLM-4.7 Flash Abliterated

Images in the input

gemma-4-26b-abliterated is the cheapest route that accepts image input, at $0.25 input and $0.75 output per 1M. Use it for screenshots and scanned documents.

Gemma 4 26B Abliterated

Longest input

glm-5.3-flash-abliterated is the cheapest route with the largest context in this set, 1M tokens. Use it for long logs, large codebases, or full reports in one request. It also has the highest output limit, 131,072 tokens per reply.

GLM-5.3 Flash Abliterated

Repeated long prompts

glm-5.3-flash-abliterated lists the lowest cached-input rate, $0.06 per 1M. Use it when the same system prompt or document goes out on every call.

GLM-5.3 Flash Abliterated

Staying on one model family

Zhipu: glm-4.7-flash-abliterated, glm-5.3-flash-abliterated, glm-5.3-abliterated. Google: gemma-4-26b-abliterated. DeepSeek: deepseek-v4.1-flash-abliterated. Pick the family your prompts were written for. The highest-priced route is glm-5.3-abliterated; price is not a measure of quality.

GLM-5.3 Abliterated

Questions

Abliterated models, in questions.

01What is an abliterated model?+

An abliterated model is an open-weight model whose refusal behaviour has been removed from the weights, so it answers some requests the original model declines. The model is otherwise the same: same context window, tool calling and output format. RouterPlex serves 5 of them over the same API as its other models: glm-4.7-flash-abliterated, gemma-4-26b-abliterated, glm-5.3-flash-abliterated, deepseek-v4.1-flash-abliterated and glm-5.3-abliterated.

02What is the difference between abliterated, uncensored and a jailbreak?+

Abliteration edits the weights directly to remove the refusal behaviour, with no further training. "Uncensored" is a looser label, often used for models fine-tuned on examples where they answer instead of refusing. On RouterPlex, gemma-4-26b-abliterated is a fine-tune of that kind and is listed with the abliterated models because it does the same job. A jailbreak is a prompt written to get a stock model past its refusals. RouterPlex does not offer or support jailbreaks, and its Terms bar circumventing the safety systems of the Service. The Terms apply to every request whatever the model.

03Are these abliterated models on OpenRouter?+

No. RouterPlex does not sell these routes through OpenRouter. When we checked OpenRouter's public model list on 10 November 2026, it had no model with "abliterated" in its name or description, and none of the RouterPlex ids. Call them on RouterPlex with an OpenAI- or Anthropic-compatible client.

04Is there a free tier for abliterated models?+

No. There is no free credit on these models. The $1 trial key gets HTTP 402 with the error type paid_model_required, and promotional credit (match bonuses, vouchers, sign-up and volume bonuses) cannot pay for them. They draw only on credit you have paid for. A card top-up starts at $5.

05Can bonus or match credit pay for abliterated models?+

No. The Terms (section 03) say promotional credit, including top-up match bonuses, vouchers, sign-up and volume bonuses, cannot be used on abliterated models; those models draw only on credit you have paid for. Promotional credit is used on other models first.

06Do you log requests to abliterated models?+

Yes. The Terms (section 05) and the Privacy Policy (section 02) say RouterPlex stores the requests you send to abliterated models, with your account identifier, the model name and the time, for security, abuse prevention and enforcement of the Terms. Stored requests are kept for at least 1 day and no more than 365 days, and are then deleted. Access is restricted to RouterPlex. Requests to all other models are not stored.

07Can I pay for abliterated models with crypto?+

Yes. A crypto top-up is paid credit, so it can pay for abliterated models. USDT, BTC, ETH and other coins are accepted, with a $15 minimum per crypto top-up. See how to pay with crypto for the steps.

08Are these the model makers' own prices?+

No. Each abliterated model bills at a RouterPlex-set rate shown in the catalog. They are not the model makers' own models or prices. RouterPlex lowers a rate when its upstream cost falls, and the table on this page shows any current cut.

Get access

Top up, then call any abliterated route.

Card top-ups from $5, crypto from $15. No subscription. The same key reaches the rest of the catalog.

Abliterated Models API: LLM Prices and Rules · RouterPlex