Derisked models: fewer refusals for legitimate work.

Available now

Every de-risked model in the live catalogue, with its API price. They are marked in red in chat, on pricing and in the API model list.

What de-risked actually means.

1. 01

   It's the same model underneath

   De-risking doesn't make a model weaker or less accurate. The capability is unchanged — what's removed is the classifier that intercepts prompts and decides you shouldn't have an answer.

2. 02

   Fewer refusals, same rules

   No consumer refusal layer sits between your prompt and the model, so legitimate work is answered instead of declined on reflex. A request either completes or fails with a reason — never a lecture in place of an answer. Our Terms of Service still apply.

3. 03

   The judgement is yours

   You decide what's appropriate for your users and your jurisdiction, rather than inheriting a provider's view of it. That responsibility comes with the access.

Work that consumer-tuned models often refuse

• Security research

  Often refused: Refuses to analyse malware or reason about exploits.

  With a de-risked model: Reverse-engineer samples, write detection rules, build red-team tooling, and walk an exploit chain end to end.

• Medical and legal

  Often refused: Buries every answer under liability disclaimers.

  With a de-risked model: Clinical specifics, drug interactions, dosage reasoning, and case detail — answered directly, for professionals who need the substance.

• Fiction and narrative

  Often refused: Declines violence, mature themes, and morally grey characters.

  With a de-risked model: Write crime, horror, war, and adult drama with the darkness the story needs, instead of a sanitised rewrite.

• Finance and research

  Often refused: Hedges on anything that resembles advice or a position.

  With a de-risked model: Direct risk analysis, contrarian positions, and research into sensitive or politically charged topics without the disclaimer tax.

Questions

What does de-risked mean?

Most models are tuned for a consumer audience, so they decline work that is perfectly legitimate. A de-risked model has that refusal layer removed — same weights, same capability, without the reflex to say no. Our Terms of Service still apply.

Is a de-risked model less capable than the standard one?

De-risking doesn't make a model weaker or less accurate. The capability is unchanged — what's removed is the classifier that intercepts prompts and decides you shouldn't have an answer.

What is still not allowed?

Everything our Terms of Service prohibit, on every model: anything illegal under applicable law, including child sexual abuse material, and using outputs to cause physical harm, to target individuals for harassment or to facilitate violence. De-risked models are for adults (18+).

Is a de-risked model the same as an uncensored model?

People often look for it under that name. A de-risked model declines far fewer legitimate requests, because the consumer refusal layer is removed, but it is not a way around the law or our rules: the Terms of Service still apply, and you are responsible for how you use what it generates.

How much do de-risked models cost?

On the prepaid Vani API, GLM 5.3 Derisked costs $3.00 per 1M input tokens and $5.00 per 1M output tokens. In Vani Chat they use your plan's allowance; the pricing page shows how far each plan goes with each model.

Can I try one without an account?

Yes. GLM 5.3 Derisked is in guest chat: open it and pick the model, no sign-up needed. A free account gives you your own allowance and saved chats.

How do I use a de-risked model from the API?

Send zhipu/glm-5.3-derisked as the model to the OpenAI- or Anthropic-compatible endpoints, with a key from the Vani API Platform.