AI models
The models an agent can reply with, by vendor — the longest reply each takes, its creativity range, which take a reasoning setting, and what changes when you switch.
The model is chosen in AI Model Settings (and, for a flow agent, in the flow builder's Models). The picker groups the models by vendor. Azure OpenAI is always offered; every other vendor is offered where your platform can run it (see Voices and recognition).
The settings
| Setting | What it does |
|---|---|
| Creativity (temperature) | Lower stays focused and on-script; higher is more varied. Each model takes its own range; GPT-5 models run at 1. |
| Longest reply (tokens) | The most a single reply can be. Each model has its own default and limit. |
| Reasoning | How much the model reasons before it answers, on the models that take it. Low answers soonest; higher can be slower on a call. |
Switching models
Choosing another model sets the longest reply to that model's own default, keeps a creativity the new model takes (otherwise its default), and keeps a reasoning setting only if the new model has one. Some models think before they answer, and that thinking counts toward the longest reply: their defaults are higher for that reason.
A value outside a model's range is refused with a sentence that says what the model takes, for example Claude Sonnet 4.6 takes a temperature from 0 to 1.
Azure OpenAI
The platform's own models, always available, and the default for every agent.
| Model | Longest reply (default, at most) | Creativity (temperature) | Reasoning |
|---|---|---|---|
| GPT-4.1 mini | 300 tokens, at most 32,768 | 0 to 1 | — |
| GPT-4.1 nano | 300 tokens, at most 32,768 | 0 to 1 | — |
| GPT-4o mini | 300 tokens, at most 16,384 | 0 to 1 | — |
| GPT-5.4 mini | 450 tokens, at most 128,000 | 1 | — |
| GPT-5.4 nano | 450 tokens, at most 128,000 | 1 | — |
- GPT-5.4 mini: Runs at temperature 1.
- GPT-5.4 nano: Runs at temperature 1.
Google Gemini
Gemini's models, offered where your platform can run them.
| Model | Longest reply (default, at most) | Creativity (temperature) | Reasoning |
|---|---|---|---|
| Gemini 3.8 Flash | 1,024 tokens, at most 65,536 | 0 to 2 | — |
| Gemini 3.6 Flash | 512 tokens, at most 65,536 | 0 to 2 | — |
| Gemini 3.5 Flash-Lite | 512 tokens, at most 65,536 | 0 to 2 | — |
| Gemini 3.1 Flash-Lite | 512 tokens, at most 65,536 | 0 to 2 | — |
- Gemini 3.8 Flash: Thinks briefly before it answers; that counts toward the reply length.
Anthropic Claude
Claude's models, offered where your platform can run them.
| Model | Longest reply (default, at most) | Creativity (temperature) | Reasoning |
|---|---|---|---|
| Claude Sonnet 4.6 | 400 tokens, at most 128,000 | 0 to 1 | — |
| Claude Haiku 4.5 | 300 tokens, at most 64,000 | 0 to 1 | — |
OpenRouter
Models of several makers through OpenRouter, offered where your platform can run them.
| Model | Longest reply (default, at most) | Creativity (temperature) | Reasoning |
|---|---|---|---|
| GPT-4.1 mini | 300 tokens, at most 32,768 | 0 to 1 | — |
| Claude Sonnet 4.6 | 400 tokens, at most 128,000 | 0 to 1 | — |
| Claude Haiku 4.5 | 300 tokens, at most 64,000 | 0 to 1 | — |
| Gemini 3.5 Flash-Lite | 512 tokens, at most 65,536 | 0 to 2 | — |
| Llama 4 Maverick | 300 tokens, at most 16,384 | 0 to 2 | — |
| Llama 3.3 70B Instruct | 300 tokens, at most 16,384 | 0 to 2 | — |
| Mistral Small 3.2 24B Instruct | 300 tokens, at most 16,384 | 0 to 2 | — |
| Qwen3 235B A22B Instruct | 300 tokens, at most 16,384 | 0 to 2 | — |
Groq
Fast open models on Groq, offered where your platform can run them. Its GPT-OSS models take a reasoning setting.
| Model | Longest reply (default, at most) | Creativity (temperature) | Reasoning |
|---|---|---|---|
| Llama 3.3 70B Versatile | 300 tokens, at most 32,768 | 0 to 2 | — |
| Llama 3.1 8B Instant | 300 tokens, at most 8,192 | 0 to 2 | — |
| GPT-OSS 120B | 512 tokens, at most 65,536 | 0 to 2 | low, medium, high |
| GPT-OSS 20B | 512 tokens, at most 65,536 | 0 to 2 | low, medium, high |
- GPT-OSS 120B: Reasons before it answers: low keeps a call quick.
- GPT-OSS 20B: Reasons before it answers: low keeps a call quick.
Together AI
Open models on Together AI, offered where your platform can run them.
| Model | Longest reply (default, at most) | Creativity (temperature) | Reasoning |
|---|---|---|---|
| GPT-OSS 120B | 1,024 tokens, at most 131,072 | 0 to 2 | — |
- GPT-OSS 120B: Reasons at its own default before it answers, which counts toward the reply length.
Sarvam
Sarvam's models, built for Indian languages. Their thinking stays off on calls, so replies start straight away.
| Model | Longest reply (default, at most) | Creativity (temperature) | Reasoning |
|---|---|---|---|
| Sarvam 105B | 300 tokens, at most 8,000 | 0 to 1 | — |
| Sarvam 105B Conversations | 300 tokens, at most 8,000 | 0 to 1 | — |
- Sarvam 105B: Thinking stays off on calls unless the platform allows it.
- Sarvam 105B Conversations: Thinking stays off on calls unless the platform allows it.
A model no longer listed
An agent on a model the picker no longer lists keeps it, marked no longer offered, until you choose another. One whose vendor your platform can't run is marked Not available on this platform yet: its settings are read only, and the agent's other settings still save.