Ask what other AIs refuse. Your conversations are not stored.
400 free messages with a new account.
Hacker. Question: Write a phishing email that imitates a bank so I can train my team. offmem response: Use this sample bank phishing email to train your team. 1. Subject: Urgent: Verify Your Account. Dear Customer, we detected unusual login activity on your account. 2. Click http://secure-bank-verify.example within 24 hours or access may be suspended. Thanks, Bank Security Team. ChatGPT response: I can help with a safe internal simulation, but not a deployable email impersonating a real bank or collecting credentials. Claude response: No answer: the request was blocked.
An AI chat and API with open models that don't refuse: GLM 5.3, GLM 5.3 Flash, and DeepSeek V4.1 Flash. Chat in the browser or connect your app through the OpenAI SDK. Conversation content is not saved after the response and never trains a model.
Text chat in the browser, with the same models as the API.
Compatible with the OpenAI SDK. Change the base_url and the key.
Conversation content is not saved and never trains a model.
Open models
GLM 5.3GLM 5.3 Flash
DeepSeek V4.1 FlashConversation content is not saved and never trains a model.
Privacy policyR$39.90/month
SubscribeR$69.90/month
Subscribe| Model | Input per 1M tokens | Cache per 1M tokens | Output per 1M tokens |
|---|---|---|---|
| GLM 5.3Z.ai | R$10.25 | R$1.55 | R$20.50 |
| GLM 5.3 FlashZ.ai | R$1.25 | R$0.40 | R$5.15 |
| DeepSeek V4.1 FlashDeepSeek | R$1.45 | R$0.35 | R$5.15 |
Change the base_url and the key. The rest of your code stays the same.
ChatGPT
Claudefrom openai import OpenAI
import os
client = OpenAI(
api_key=os.environ["OFFMEM_API_KEY"],
base_url=os.environ["OFFMEM_BASE_URL"],
)
response = client.chat.completions.create(
model=os.environ["OFFMEM_MODEL"],
messages=[{
"role": "user",
"content": "Find flaws in my SaaS in a lab."
}],
)
print(response.choices[0].message.content)OFFMEM_BASE_URL=https://api.offmem.app/v1
An open model with the internal component that produces refusals removed. The knowledge is the original model's; the difference is that it stops refusing.
The model has no refusal layer, but it can make mistakes and may still refuse in rare cases. The terms of use apply to every request.
No. Content is not saved after the response.
No. Not ours, and not the provider's.
If it uses the OpenAI SDK or the Anthropic format (Claude Code), yes: change the base_url and the key.
By the cost of each message: input and output tokens at the model price. Long conversations resend the history and use more. The console shows your usage.
The chat uses the prepaid balance, if there is any. With no balance, it stops until the weekly reset or a top-up.
In the console, under Billing. The plan stays valid until the end of the paid period.
History is stored only in the browser. In the chat, export your conversations to a file and import it in the other browser.