Skip to main content
Reasoning effort controls how much reasoning a supported model uses. Set a saved default on an assistant, override it for a conversation, or choose a different level for one message. The selected model determines which levels are available.

Find compatible levels

Request the current catalog with your API key:
The response’s data array groups models by provider. Each model includes its identifier, reasoningEfforts, and reasoning/vision capabilities. A provider’s reasoningEfforts is the union of its selectable models’ levels. Use the model’s list to choose a value. An empty list means no named effort can be selected; omit the field or use null to keep the model’s default behavior. The API recognizes none, minimal, low, medium, high, xhigh and max. Each model supports only a subset. Unsupported combinations return HTTP 400, including when changing the model while retaining an incompatible effort.

Assistant default

Include reasoningEffort when creating an assistant or updating it:
The CLI’s assistants create command accepts the same flag, including alongside --from-json. Set null in API JSON, or --reasoning-effort default in the CLI, to remove the assistant’s explicit effort and use the adapter default.

Conversation and message overrides

Send these fields to the message endpoint:
This turn uses high; following turns use the saved conversation value low unless they include another message override. The message-level setting also applies to tool iterations and resumptions within that turn. Both synchronous and asynchronous requests accept these fields. Omit chatUid to start a new conversation with the same settings.
Resolution is message → conversation → assistant → adapter default. none is an explicit model level, distinct from inheritance. The CLI translates default (also accepted as null) into JSON null. API JSON must use null, not the string "default". Explicit CLI flags override JSON input, including when clearing a setting. To restore assistant inheritance for an existing conversation:

Read the configured and effective values

Assistant reads include reasoningEffort. Conversation history includes reasoningEffort (the saved conversation setting) and effectiveReasoningEffort (the last recorded model call). Individual model response messages expose messageTokenUsage.reasoningEffort. An absent effective value indicates that the adapter’s existing default behavior was used.
In the console, the effort selector appears next to the model selector. Workspace assistant settings also provide it. Linked workspace assistants inherit the source setting unless a local override is saved; clearing that override restores inheritance. Workspace configuration uses its existing session-authenticated API and workspace permissions.