---
title: "Choose the AI model behind your agent — Dazza help"
description: "Pick the model that writes Dazza’s replies: included Claude Sonnet, Opus or Haiku, or your own Anthropic, OpenAI or OpenAI-compatible key for Llama or Mistral."
url: https://dazza.io/help/ai-model
updated: 2026-10-26
---

# Choose the AI model

Use the included Claude models, or your own Anthropic, OpenAI or OpenAI-compatible key, for Llama, Mistral or your own server.

*1 min read · Updated 26 Oct 2026 · 5 sections*

The AI model writes Dazza’s replies, in the preview and to customers. It also answers setup’s test questions. Choose it in **Settings → AI model**.

## Included Claude

Nothing to set up. Choose one of:

| Model | What it’s like |
| --- | --- |
| Claude Sonnet 5.5 | Balanced. The default. |
| Claude Opus 5.5 | The most capable, a little slower. |
| Claude Haiku 5.5 | The fastest. |

## Your own key

- **Your Anthropic key.** Claude, with your own Anthropic account.
- **OpenAI.** Your OpenAI key and a model name, such as `gpt-5.6-terra`. Type the name exactly as OpenAI writes it.
- **Llama, Mistral or your own.** Any endpoint that works like OpenAI’s. Enter its base address, your key and the model name. For example:
  - `https://api.groq.com/openai/v1`
  - `https://api.together.xyz/v1`
  - `https://openrouter.ai/api/v1`
  - or your own server

With your own key, conversations are processed under your agreement with that provider, and they bill you directly. There’s no model cost in Dazza’s usage figures.

Keys are encrypted and never shown again. Leave the key empty to keep the saved one.

## Check it works

**Check the saved model** sends one message, with no customer data, and shows how long the reply took.

## If it stops working

If the model stops working, Dazza hands customers to your team instead of guessing. **Settings → AI model** shows the last problem.

## Workflows and tools

Workflows and [MCP tools](https://dazza.io/help/connections) work with every option.