> ## Documentation Index
> Fetch the complete documentation index at: https://docs.rovax.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Model Selection

> How to choose the provider and LLM model for each agent in Timely.ai, capability comparison, credit cost, and selection best practices.

The language model defines the intelligence, speed, and cost of each agent execution. Timely.ai supports four providers with access to multiple models — each agent can use a different model.

## Where to select the model

<Steps>
  <Step title="Access the agent settings">
    In the agent list, click the agent you want to configure.
  </Step>

  <Step title="Open the Model tab">
    Click the **Model** tab in the agent settings. The configuration dialog displays three columns: providers, available models, and behavior parameters.
  </Step>

  <Step title="Select the provider">
    In the left column, choose the provider: OpenAI, Anthropic, Google, or OpenRouter. The center column updates with the available models for the selected provider.
  </Step>

  <Step title="Select the model and save">
    Click the desired model in the center list. Adjust temperature and effort in the bottom section and click **Save settings**. The change takes effect on the next execution.
  </Step>
</Steps>

## How providers work

Timely.ai offers two modes of model access: using the platform's credits (access via Timely) or connecting your own API key from an external provider.

### Option 1 — Using Timely credits

Each model consumes a fixed number of credits per execution. You do not need to create accounts with providers — Timely manages authentication and billing.

<Steps>
  <Step title="Select any available model">
    All models listed in the interface are immediately available if you have credits in the workspace.
  </Step>

  <Step title="Check the credit cost">
    Each model displays the estimated cost per execution. Actual consumption varies based on prompt size, conversation history, and the number of tool calls in the same execution.
  </Step>

  <Step title="Monitor consumption">
    Track credit consumption in **Settings > Billing** to adjust the model as conversation volume grows.
  </Step>

  <Step title="Reload credits when needed">
    Credits can be reloaded at any time from the billing panel. Agents without available credits automatically pause their responses.
  </Step>
</Steps>

### Option 2 — Connecting your own provider

Workspaces on the Enterprise plan can configure custom providers, connecting models with their own API keys or open-source models hosted on internal infrastructure.

<Steps>
  <Step title="Access workspace settings">
    Go to **Workspace Settings > AI Models** in the administration menu.
  </Step>

  <Step title="Add the provider credentials">
    Enter the provider's API key (OpenAI, Anthropic, or Google). The key is stored with encryption and used exclusively for your workspace's executions.
  </Step>

  <Step title="Select the custom provider in the agent">
    When configuring the agent model, the custom provider appears as an additional option in the providers column.
  </Step>

  <Step title="Confirm the connection with a test">
    Use the Internal chat to send a test message and confirm the custom provider is responding correctly before publishing the agent to production.
  </Step>
</Steps>

<Note>
  Custom provider configuration requires an Enterprise plan. Contact support to enable this feature in your workspace.
</Note>

## Switching model per agent

Each agent can use a different model within the same workspace. This allows you to optimize cost and performance per use case:

* High-volume FAQ agents use cheaper models (Kimi K2.5 or Gemini 2.5 Flash)
* Complex analysis agents use high-capacity models (Claude Opus 4.6 or Gemini 3.1 Pro)
* General support agents use balanced models (Claude Sonnet 4.6, marked as recommended)
* OpenAI GPT-5.x models disable temperature — the slider becomes inactive automatically
* Anthropic models with active reasoning force temperature to 1.0 — the temperature control has no effect

## Best practices

* Always start with **Claude Sonnet 4.6** — it is the platform's recommended model for most use cases, with a good balance between quality (4 credits/execution), speed, and tool support
* Test any model change in the Internal chat with representative questions before applying to production
* Monitor latency after switching to larger models — models such as Claude Opus 4.6 and Gemini 3.1 Pro have higher latency that may affect the experience on messaging channels
* Do not switch models based on cost alone — compare response quality in the real scenarios of your use case before migrating