> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.apologist.ai/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.apologist.ai/_mcp/server.
# Choosing an AI Model
> Choose a model for an Apologist AI Agent, tune supported parameters, and verify the result
The model is the engine that turns an Agent's instructions and context into a response. Changing it can affect answer quality, instruction-following, response length, and the number of credits used for every response.
## Purpose
Choose the lowest-cost model that reliably passes the questions your Agent needs to answer, then tune only the parameters that solve a specific, observed problem.
## Prerequisites
* An active Agent with baseline instructions. See [Customizing Instructions](/console/configuration/customizing-instructions).
* A small test set of real questions, including straightforward, ambiguous, and difficult examples.
* Permission to configure models on your team's plan.
These settings belong to the Agent's default Responder, the AI configuration used to answer prompts. If your Agent has the Responders capability turned on, the **Behavior** page is hidden and you configure the model on each Responder instead.
## How the Setting Reaches the Agent
For each prompt, the Agent runtime resolves the Responder, loads its selected model, and sends only the parameters that model supports. The published Agent interface does not display the model name, so verify a change through representative questions rather than looking for a model badge in the chat window.
Model availability is live configuration. Names, recommendations, credit costs, and deprecation dates can change, so use the cards currently shown in Console rather than relying on a fixed list in documentation.
## Steps
#### Open the Model Tab
Select **Agents** in the left-hand navigation, open your Agent, choose **Behavior** from the row of pages across the top, and then select the **Model** tab.
Console groups models into three classes:
| Class | Console description | Good first use case |
| ------------ | ------------------------------------------------------------- | ---------------------------------------------------------------------- |
| **Limited** | Inexpensive models with very limited reasoning abilities | Simple, tightly constrained answers where cost matters most |
| **Standard** | Economic models that balance lower cost with better responses | General-purpose Agents and the usual starting point for comparison |
| **Premium** | Higher quality output at a premium price point | Nuanced instructions, difficult questions, and higher-stakes responses |
Selecting a class filters the model cards beneath it. Scroll within the page to see the full list for that class.
#### Choose a Model By Evidence And Cost
Each model card shows its credit cost per response. Additional feature charges, such as Live Web Search, are separate and can stack on top of this amount.
Use the card signals as a shortlist:
* **Recommended** cards are starred and sorted toward the top.
* A **Deprecation** label identifies a model that should be replaced; when available, the label includes the planned date.
* An unlabelled card is still available, but Console is not specifically recommending it.
Start with the least expensive recommended model in the class that fits your use case. Do not assume a premium model is automatically the right choice: a cheaper model that consistently follows your instructions is the better production model.
Selecting a different model resets **Max Output Tokens** on the Advanced tab to that model's maximum. Review the Advanced tab before saving, especially if your Agent should give short answers.
#### Review the Supported Advanced Parameters
Still on the **Behavior** page, select the **Advanced** tab.
Console hides parameters the selected model does not support. A missing control is therefore expected and does not mean the form failed to load.
| Setting | What it controls | Practical guidance |
| --------------------- | ----------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Reasoning Effort** | How much supported reasoning models deliberate before answering | Use **Low** for direct questions and compare **Medium** or **High** on genuinely multi-step tasks. **Minimal** is available only where the model supports it. |
| **Verbosity** | The amount of detail in supported model responses | Choose **Low**, **Medium**, or **High** only when prompt instructions alone do not produce the desired length. This field is hidden for models that do not support it. |
| **Temperature** | Response variability, from 0 to 2 | Lower it for more consistent wording; raise it only when the Agent needs more variety. |
| **Top P** | How broadly the model samples likely next tokens, from 0 to 1 | Lower values narrow the candidate set. Change this or Temperature, not both at once, so the effect remains measurable. |
| **Frequency Penalty** | Repeated use of the same tokens, from -2 to 2 | Positive values discourage repetition. Leave it at 0 unless repetition is a demonstrated problem. |
| **Presence Penalty** | Reuse of topics already mentioned, from -2 to 2 | Positive values encourage the response to introduce new topics. Leave it at 0 for focused answers. |
| **Max Output Tokens** | The maximum response-token budget | Reduce it to control overly long responses. The allowed maximum comes from the selected model. |
| **Cache Duration** | How long the same request and configuration may reuse a stored response | Leave it unselected while comparing repeated runs. Set it only when reuse is appropriate for the Agent's content. |
If **Reasoning Effort** or **Verbosity** shows **Select an option**, the Agent runtime falls back to **Medium** for a model that supports that parameter. Select an explicit value when you want the saved configuration to be unambiguous.
Change one parameter at a time and retest. Otherwise, you cannot tell which setting improved or degraded the response.
#### Save And Test On the Published Agent
Select **Save Changes**. Model and Advanced are tabs in the same Behavior form, so one save persists changes made on either tab.
Open the Agent's published URL. In the standalone interface, enter a question in **Ask a question here**, select **Submit**, and use **New Chat** before each isolated test so conversation history does not affect the next result.
Run the same test set for every candidate model:
| Test | What to compare |
| ----------------------------------------- | --------------------------------------------------------------------------------- |
| A straightforward, source-backed question | Factual accuracy, citation quality, and instruction-following |
| An ambiguous question | Whether the Agent asks for clarification instead of guessing |
| A multi-step or nuanced question | Reasoning quality and whether important constraints are missed |
| A requested short and long answer | Response length, completeness, and whether Max Output Tokens truncates the result |
Record the model, parameter values, credits per response, and pass/fail result. Judge the complete set rather than the most impressive single answer; model output can vary between runs.
A configured **Cache Duration** can return the same stored response for repeated identical messages under the same model configuration. Leave it unselected while measuring response variation.
Test through the published Agent when validating Console settings. API clients can supply per-request model settings that override the saved values and are therefore a different test.
## Expected Result
The Agent uses a non-deprecated model that passes its representative test set at an acceptable credit cost. Advanced parameters are explicit where needed, unsupported controls remain hidden, and the published Agent responds consistently with its instructions.
## Troubleshooting
* **The Behavior page is missing.** The Agent has the Responders capability turned on, so choose and configure the model on each Responder.
* **The Model or Advanced controls are disabled.** Model configuration is not included in the team's current plan, or editing is restricted for this Agent.
* **A model is not listed.** Console shows active models allowed by the team's plan. Availability changes as models and plans change.
* **An Advanced control disappeared after switching models.** The new model does not support that parameter. Console intentionally hides it, and the Agent runtime does not send it.
* **Max Output Tokens changed unexpectedly.** Choosing a model resets the value to that model's maximum. Set the intended limit after selecting the model.
* **Reasoning Effort says Select an option.** With no explicit value saved, the runtime uses Medium on supported reasoning models.
* **Repeated tests return exactly the same answer.** Clear **Cache Duration** and save, then start a New Chat before testing again.
* **The response still seems to use the old behavior.** Confirm **Save Changes** completed, start a New Chat, and test through the published Agent rather than an API request carrying overrides.