Adaptive thinking, extended thinking and effort

Three names, three different things, and the documentation uses all of them.

Extended thinking — the old, explicit control

thinking.type: "enabled" on the request. You turn reasoning on; you can budget it. Supported on Claude Haiku 4.5, Sonnet 4.5 and Opus 4.5. Listed as deprecated on Opus 4.6 and Sonnet 4.6. Not supported at all on any Claude 5 model.

Adaptive thinking — the current model

The model decides how much to think. There is no parameter to enable it.

Model	Extended thinking parameter	Adaptive thinking
Claude Fable 5	No	Yes, always on
Claude Opus 5	No	Yes
Claude Sonnet 5	No	Yes
Claude Haiku 4.5	Yes	No
Claude Opus 4.8 / 4.7	No	Yes
Claude Opus 4.6 / Sonnet 4.6	Yes (deprecated)	Yes
Claude Sonnet 4.5 / Opus 4.5	Yes	No

Haiku 4.5 is the only current model on the old mechanism. Any code that must run on both Haiku 4.5 and a Claude 5 model needs a branch.

Effort — the dial you still control

The effort parameter is how much of that adaptive reasoning you want.

• Opus 5 and Sonnet 5: defaults to high on the Claude API and Claude Code.
• Opus 4.8: defaults to high on all surfaces, including claude.ai.

Set it explicitly to use a different level. Since thinking tokens are output tokens, and output is five times the price of input on every current model, effort is one of the larger cost levers available — and one of the most commonly left at default.

Not to be confused with fast mode

Fast mode changes the speed of generation at premium price. Effort changes how much generation happens. They are orthogonal, and fast mode does not reduce thinking.