Model comparison

Every AI agent runs on a language model (LLM) that you choose in its LLM Configuration section. The model affects how fast the agent replies, how well it follows long instructions, and how much each reply uses of your plan.

This page covers models for chat and email agents. For voice agents, see voice LLM settings.

Available models

The AI Model dropdown lists the models below. Which models you can use depends on your plan and usage. Models not included are shown but disabled, with a prompt to upgrade.

Choose a starting model

Start with a fast, low-cost model: Gemini Flash 2.5 or Gemini Flash 3.0. They reply quickly and handle most support conversations well. Move an agent to a larger model only when testing shows it needs one.

Larger models are usually slower and use more of your plan per reply. See usage to track what your agents use.

The default model

A new agent opens with OpenAI GPT 4.1 Mini selected, a temperature of 0 and 4,000 max tokens. Change it under LLM Configuration before you publish if you want a different model. An agent with no model saved at all runs on Gemini Flash 2.5.

When a model is unavailable

If the chosen model fails to respond, Robylon retries the request on a backup model so the customer still gets a reply. For example, Gemini Flash 2.5 falls back to a newer Gemini Flash model, and other models fall back to a comparable model from another provider.

How temperature and tokens affect replies

These settings sit next to the model in LLM Configuration:

• Temperature: keep it at or near Deterministic (0) for support, so the same question gets the same answer.
• Max Tokens: the longest reply allowed, from 10 to 8,000. The default is 4,000. Lower it if replies run long.
• Thinking Tokens: how much the model can reason before replying, from 0 to 8,000. The default is 0 (off). Raise it for agents that make multi-step decisions, and expect slower replies.