Token Economics

Adaptive Thinking

The AI decides for itself how long to think about each request, guided by how much effort you ask for.

Definition

A thinking mode (thinking: {type: "adaptive"}) where the model decides when and how much to reason, steered by effort. It replaces fixed budget_tokens thinking budgets, which return a 400 error on Claude Opus 4.7 and later and on 5.x models.

Example

With thinking: {type: "adaptive"}, a typo fix gets almost no thinking while a race-condition bug gets thousands of thinking tokens.