---
title: LLM configuration
description: "Which language model answers for an agent: the Model card on text agents, Model controls on voice agents, and how the list of models is built."
---

A text agent has a **Model** card on **Settings** where you pick the LLM that
writes its replies. A voice agent keeps its model behind **Model controls**, a
button in the header of the **General** card, together with the speech services
that carry it.

:::note[When to change the model]
The default suits most agents. Change it when replies are too slow for the channel,
when a task needs stronger reasoning, or when your company standardises on one
provider.
:::

## Text agents

1. Open the text agent and stay on **Settings**.
2. Scroll to **Model**. The card reads **Pick the LLM that powers this text agent**.
3. Open the picker. Models are grouped under their provider; type in **Search
   models…** to filter.
4. Pick one and **Save**.

![The Model picker of a text agent, grouped by provider.](/media/build/text-agent-model-picker.webp)

A new text agent starts on OpenAI's **GPT-5.6 Luna**.

## Voice agents

Choose the **Model controls** button in the header of the **General** card on
**Settings**. The panel shows the current model at the top, then:

| Field | What it sets |
|---|---|
| **LLM provider**, **LLM model** | The model that writes the agent's replies |
| **Speech-to-text provider**, **Speech-to-text model** | The service that transcribes the caller. **Provider default** leaves the model to the provider |
| **Text-to-speech provider**, **Text-to-speech model** | The service that speaks the agent's replies. **Provider default** leaves the model to the provider |
| **Edge judge** | Flow agents only. The model that decides which edge to follow |
| **Variable extractor** | Flow agents only. The model that extracts variables from what the caller said |

Some models listen and speak themselves. With one of them selected, the speech
fields are disabled and read **This model listens and speaks itself, so it uses no
separate speech services**.

The model also decides which voices exist. If the new model cannot speak the
selected voice, the voice is switched for one it can, and switching the model back
restores it. A text-to-speech model that does not support the agent's language is
marked **Doesn't support this language** and cannot be picked. See
[Choosing a voice](/build/voice-and-audio/choosing-a-voice).

## What is offered

The list of providers and models comes from Callab's catalogue, not from the app,
and each model is offered for text agents, voice agents or both. Today a text agent
is offered OpenAI models: **GPT-5.4 Mini**, **GPT-5.4 Nano**, **GPT-5.5**, **GPT-5.6
Luna** and **GPT-6 Astra**. Named variants such as **Luna** and **Astra** are
distinct models, not aliases.

The catalogue changes as providers release and retire models; the names above are
not a promise that a given version stays available. An agent that holds a model the
catalogue no longer lists keeps it, and the picker still shows it as selected.

## Behaviour and limits

- The model applies to every turn of every conversation with the agent, in the
  tests, the widget and every deployment.
- Each conversation under **Conversations** shows the model that answered and the
  tokens used per turn (**Input**, **Output**, **Avg / turn**), which is how you
  compare two models on the same prompt.
- Changing the model is a saved change like any other: it creates a version and can
  be rolled back.

## Verify

On a text agent, open the agent under **Testing**, **Chat**, send a message, then open the thread
under **Conversations**. The model id is printed under the agent's name and after
each reply.

## Related

- [Choosing a modality](/build/choosing-a-modality)
- [Quickstart: your first text agent](/get-started/quickstart-text-agent)
- [Conversations](/operate/conversations)
- [Prompting guide](/build/prompting/prompting-guide)