LLM
Pick the AI model behind a voice agent, see which models are available, and set temperature for fast, accurate calls.
Set it in the voice agent under LLM Configuration. On a phone call the model has to be fast: every extra second before the agent answers is dead air for the caller.
Which model should I start with?
| Your calls | Start with | Why |
|---|---|---|
| Most use cases: sales, collections, support, multilingual calls | GPT 4o Mini | Good multilingual understanding and acceptable latency. |
| Highly scripted, high-volume calls | Gemini Flash (a Lite version if possible) | Very low latency. The risk of made-up answers is kept low by a strict script. |
| Complex calls that need careful reasoning | A larger model such as GPT 4.1 | More capable, but test the latency on real calls before going live. |
Test your choice with real call scripts. The right model is the fastest one that still handles your hardest calls correctly.
All models in the AI Model list
Choose from models by OpenAI, Google, Anthropic and several open models hosted on Azure:
| Provider | Models (as named in the AI Model list) |
|---|---|
| OpenAI | OpenAI GPT 4o, OpenAI GPT 4o Mini, OpenAI GPT 4.1, OpenAI GPT 4.1 Mini, OpenAI GPT O3 Mini, OpenAI GPT 5, OpenAI GPT 5 Mini, OpenAI GPT 5 Nano, OpenAI GPT 5.1, OpenAI GPT 5.6 Luna |
| Gemini Pro 3.0, Gemini Pro 2.5, Gemini Flash 3.6, Gemini Flash 3.5, Gemini Flash 3.5 Lite, Gemini Flash 3.1 Lite, Gemini Flash 3.0, Gemini Flash 2.5, Gemini Flash 2.5 Lite, Gemini Flash 2.0 | |
| Anthropic | Anthropic Claude Sonnet 4.5, Anthropic Claude Haiku 4.5 |
| Azure (open models) | Azure Kimi, Azure Deepseek Pro, Azure Deepseek Flash, Azure GLM, Azure GPT OSS 120B |
For a wider comparison of models across channels, see Model comparison.
How do the model families behave on calls?
- Gemini Flash models have the lowest latency. They are less capable in open-ended conversations and more likely to make things up when a call goes off script. The Lite versions are the fastest.
- Gemini Pro models are more capable but noticeably slower, which callers can hear as pauses.
- OpenAI models sit in between: a balance of capability and latency. The Mini and Nano versions are tuned for speed.
- Anthropic models are the most precise at following detailed instructions, at a higher latency.
What does temperature do?
Temperature controls how predictable the agent’s replies are, on a scale from 0 (Deterministic) to 1 (Random). The default is 0.
- Low (0 to 0.5): consistent, factual replies. Lower risk of made-up answers, but less expressive. Use this for most support, collections and confirmation calls.
- High (0.7 to 1): more varied, engaging replies, with a higher risk of made-up or inconsistent answers.
Max Tokens and Thinking Tokens, also in LLM Configuration, have no effect on voice calls; to keep replies short, ask for short answers in the prompt.