Adaptive thinking

How it works

Adaptive thinking means the model itself decides whether and how much to reason on each request, instead of following a fixed thinking budget you set. Easy prompts may get little or no reasoning; hard ones get more.

Makers that offer it usually let you steer it with an effort setting and report reasoning tokens in the usage data. On some models it is always on; others offer it next to a fixed-budget or no-thinking mode. The details are maker-specific, so check the model's docs.

It simplifies mixed workloads where difficulty varies, but makes per-request cost and latency less predictable. Watch the spread of reasoning tokens across requests, not just the average.

Worked example

Illustrative: in a support bot, a password-reset question uses no reasoning tokens while a billing dispute spanning three invoices uses 5,000. At $10 per 1M output tokens that is $0 versus $0.05 of reasoning per request, decided by the model rather than your code.

Common mistakes

• 1Assuming cost per request is stable when reasoning varies with difficulty.
• 2Planning latency from typical requests when the slow tail comes from hard ones.
• 3Not lowering effort when the model over-thinks routine requests.