Agents

Choose a model and reasoning effort

Pick the model that answers for an agent, set how hard it reasons before it replies, and understand the latency and cost that comes with each choice.

Every agent runs on one model, and on models that support it you can also set how hard that model reasons before it answers. Both live on the Voice tab, under Model, described as The provider and model that power replies.

What you need

  • An agent that is not archived. The Change button is disabled on an archived agent.
  • Nothing else. The model list is offered to every workspace member who can open the agent.

Steps

Step 1: Open the Model section

Open the agent and select the Voice tab. Scroll past Used Voices to Model. The card shows the provider name above the model id, for example OpenAI above gpt-5.4.

Step 2: Open the picker

Press Change. A sheet titled Model opens with the current model and, on models that accept it, the reasoning control. Press the model row to open Choose a model.

Step 3: Pick a model

The list is grouped by provider, groups sorted alphabetically. A model still in preview carries a Preview chip. Press one to select it and the picker closes on the new choice. Nothing is saved yet.

Under each model's name the picker shows ≈ N credits / min on your plan: an estimate of a voice-call minute on that model, 2,500 credits for the call plus about 16,000 AI tokens, at the rate of your plan's credits. Hover the line for the provider's price per 1M tokens, the 10% fee, and the same prices in credits on your plan.

You never choose the provider. It follows the model you pick, which is what the hint under the list means: The model this agent runs on. The provider is set automatically.

Step 4: Set the reasoning effort

The Reasoning effort control appears only when the selected model supports it. Its five segments are Default, Minimal, Low, Medium and High, and the portal describes them as How hard the model reasons before it answers. Higher effort can improve quality but adds latency.

Default sends no reasoning setting at all, so the model does whatever it does on its own.

Step 5: Close the sheet

Press Done. Closing the sheet saves the agent.

Get it right

Reasoning happens before the first word, so on a voice call it is heard as silence while the caller waits. Match the effort to how the caller experiences a pause.

Recommended

  • Voice calls: Default or Minimal, and let the prompt do the work of keeping answers short.
  • Helpdesk tickets and email-shaped replies: Medium or High, where a second of thinking costs the reader nothing.

Less effective

  • High on a live phone agent, which buys a better sentence at the price of a silence the caller hears as a dropped call.
  • Raising the effort to fix an answer that is wrong because the prompt is vague or the document is missing.

Worked example

A support agent on the phone is answering well but slowly. You open Model, leave the model as it is and move Reasoning effort from High to Minimal, then press Done.

The next call starts on the new setting; the one in progress does not change. Compare the model's time on a few turns in Conversations before and after, and check the credits on those turns: less reasoning means fewer tokens, so the model's share of each turn falls with the latency.

What happens next

The next conversation starts on the new model and effort. A call already running finishes on the model it started with.

Which models you can pick

The picker is not a fixed list. It shows the chat models available to your workspace, so it changes as they change:

  • A model that is retired stops appearing. If an agent is still set to it, calls use the default model, and the page shows a red note naming it until you pick another model.
  • A newly available model can take up to 5 minutes to appear in the picker, and up to 24 hours before its reasoning support is recognised.

Which models accept reasoning effort

The Reasoning effort control appears only on models that accept it, which today are the o1, o3, o4 and gpt-5 families.

The same rule is applied when you save. Move an agent to a model that does not accept the setting and the effort is cleared and stored empty, so a stale effort is never sent to a model that would reject it.

What reasoning effort changes

Reasoning effort is passed to the model unchanged, so the exact behaviour is the model's own. The trade it makes is always the same shape.

Setting What to expect Where it fits
Default Whatever the model does unprompted Start here and change it only if you have a reason
Minimal The least deliberation, the fastest first word Voice calls where a pause is worse than a slightly worse answer
Low A small amount of deliberation Short factual exchanges
Medium Noticeably more deliberation Text channels where the customer is reading, not waiting on a line
High The most deliberation, the longest wait Tickets and email-shaped work, rarely live voice

Two costs move together with the setting.

Latency. Reasoning happens before the first token, so on a voice call it lands as silence between the caller finishing and the agent starting. The model's time is recorded on every turn, so you can compare settings against real conversations rather than guessing.

Credits. Every turn reports its prompt and completion tokens, and the turn's AI usage is priced from them at the provider's price plus 10%. More reasoning means more tokens, so a higher effort costs more per turn on the same conversation.

Temperature

Every agent runs at temperature 0.7. Changing it is not available from the portal today.

Limits

Limit Value Applies to
Reasoning effort values Empty, minimal, low, medium or high The Reasoning effort control
Reasoning effort on an unsupported model Cleared on save The Reasoning effort control
Temperature 0.7 Every agent
Models offered The chat models available to your workspace The picker
A new model appearing in the picker Up to 5 minutes The picker
Reasoning support recognised on a new model Up to 24 hours The Reasoning effort control
When a saved change applies The next conversation The agent

Troubleshooting

This model is no longer available. Pick another to keep this agent working.

The model saved on this agent is not in the current list. The agent keeps starting conversations on it, and nothing guarantees it will answer. Pick a current model. Until you do, you can still save changes to the agent's other settings.

Couldn't load the model list. Showing your current selection.

The list could not be fetched. The picker falls back to a single row holding the agent's current model, so you can close the sheet without changing anything. Reopen it later.

The reasoning control is not in the sheet

The selected model does not accept a reasoning setting, or the list failed to load and the fallback row reports none. Pick a supported model and the control appears.

FAQ

What does Default actually do?

It sends no reasoning setting, and the model behaves as it does without one.

Does a higher effort cost more?

Yes. Reasoning produces tokens, tokens are reported on the turn, and the model's share of the turn's credits is priced from them.

Can I set a different model per channel?

No. The model belongs to the agent, and every channel it serves uses it.

Can I change the temperature?

Not from the portal today. Every agent runs at 0.7.

Does changing the model interrupt a call in progress?

No. The change applies at the start of the next conversation.

Which provider will my agent use?

The one that offers the model you picked. The picker groups models by provider, and the provider is always the one the chosen model belongs to.