Thinking Directive Modes

This reference lists the Thinking directive modes on the Tier map of a PaletteAI Inference Launchpad appliance, the value each mode carries at request time, the console states that override the setting, and how each engine family interprets a token budget. For the concept behind the directive, refer to The Thinking Directive. For the steps to set the directive, refer to Set the Thinking Directive for a Tier.

Modes

Mode	Authored value	What the model does
off	off	The model does not reason. It answers each request directly.
on	on	The model reasons at the depth its own defaults choose. No token budget is set.
budget	budget:<n>	The model reasons with a token-budget hint of <n> tokens. The value is a hint, not a hard cap.

Per-Engine Interpretation of Budget

Engine family	How the value is treated
Anthropic frontier, version 4.6 and earlier	A target the provider aims for, not a limit.
Anthropic frontier, version 5 and later	No observable effect. Reasoning runs to the full output allowance.
Local models	Not read. No chat template that ships with the appliance reads the value on any family tested.

Console States That Override the Setting

State	When it appears	Effect
off is disabled	The tier resolves to an OpenAI o-series or GPT-5 family model.	The selector does not offer off for that tier.
Grey chip reading ignored — no thinking surface	The tier's resolved model runs on an engine that has no reasoning surface.	The engine ignores the directive; the request is served normally.