Language models

Choose the model that writes your agent's replies, set its temperature, and understand speed, language fit and cost.

What you will achieve

You will pick a model that balances reply speed, quality in your language and price.

Step-by-step setup

Open the agent and the Assistant tab.
Click the Model card. The Model panel opens with the note: "The model decides what the agent says and most of its response delay."
Open Language model. The list shows each model's name and its price per million tokens in and out. Type in the list's search to filter it.
Choose a model. For models that allow it, set Temperature (see Conversation settings).
Press Save, then test with Talk.

The models

OpenAI: the GPT-4o and GPT-4.1 families (including the mini and nano sizes), the GPT-5 and GPT-5.x families, and the GPT-6 and 6.1 families (Astra, Sol, Luna).
Google: Gemini 3.1, 3.5, 3.6, 3.7 and 3.8 Flash and Flash-Lite models.

The picker is the authority on what is offered today and at what price, so use it rather than a list in a guide. It also marks when a price is due to change.

Configuration details

Reasoning models and Gemini models set their own temperature, so no slider is shown. Hidden "thinking" they do is billed as output tokens and can add delay.
Language fit. For languages other than English the picker rates each model great, ok or weak. Small models lose a lot of accuracy outside English; the panel warns ("This model is weak in Hindi") and offers a stronger model with one click.
Speed. Larger and reasoning models answer more slowly. For phone conversations, a fast mid-size model usually feels best. Test the delay yourself with Talk.
Cost. The Cost panel below the picker shows the per-minute estimate for your choices. See Costs.
Not in the price list. If a model name is not in the list, it is priced as GPT-4.1 mini.

Testing

Ask the same three questions on two models. Note which answers are correct and how long each takes to start speaking.

Expected result

The Model card shows the chosen model and its price per million tokens, and the Cost card updates.

Troubleshooting

Replies are slow. Try a smaller or non-reasoning model, and keep the system prompt short.
Wrong or odd answers in your language. Check the language fit warning and move up a size.
A Gemini notice says no key. See Voice and AI overview.

Something here does not match what you see in the dashboard? Let us know.