Extended thinking
How the HQ sets the thinking level, and how you set the effort for one conversation.
Some AI models think before they answer or call a tool. The HQ sets the thinking level for each request. You can raise or lower it for one conversation with the effort control.
Default effort
When you do not set an effort, the HQ asks for thinking off. Then it changes this request to a level that the model supports:
| Model type | What the HQ does |
|---|---|
| Thinking is optional | Thinking stays off. |
| The model must think | The HQ uses the lowest thinking level that the model supports. |
This is the Default choice in the effort control. Background tasks always use this rule for their model. These tasks include chat titles, the memory sweep and the skill review. Sub-agents also use this rule.
Set the effort for a conversation
The effort control is in the conversation screen of Desktop, the web app, the iOS app and the Android app. It shows Default and the levels that the model supports: Low, Medium, High and Extra high. The control shows only for models where the HQ can send a level to the provider.
Choose a level. The HQ saves it for this conversation only. The next reply uses it. Other conversations of the same squad member keep their own setting. A reply that is in progress does not change.
Choose Default to go back to the rule above. The control describes Default as The model's usual effort.
If you change the model of the squad member, the saved level stays. If the new model does not support that level, the control shows the next higher level that the model supports. If there is no higher level, it shows the next lower level. Desktop and the web app show the saved level in a tooltip. The Android app shows it below the control. Choose any level to save it again.
Thinking text in chat
Some providers send the thinking text with the answer. If the provider sends it, the client can show it. For example, the web app shows a Show thinking control on the answer.
Response time and cost
A higher effort can make a response slower and use more output tokens. Thinking can use output tokens, also when the provider sends no thinking text. A lower effort decreases the thinking effort. It does not give a fixed response time or a fixed token count. A model that must think cannot run with thinking off.
Provider plugins
If you add models through a provider plugin, the plugin catalog must tell which thinking levels the provider accepts. Set reasoning and thinkingLevelMap on each model. Use "off": null to tell the HQ that the model must think. The effort control shows the levels from this map. Refer to the example plugin for the model metadata.