Adaptive Thinking & Effort
Let Claude reason before it answers: adaptive thinking, the five effort levels, why budget_tokens is gone, summarized vs omitted thinking blocks, and preserving them across tool calls.
Last updated
After this section you can
- Turn on adaptive thinking and pick an effort level per route rather than a token budget
- Explain why budget_tokens is gone and what replaced it
- Preserve thinking blocks across tool calls, and know why the raw reasoning is never returned
Adaptive Thinking & Effort: Reasoning on Demand
Some problems need more than a fast answer — they need the model to work through them. Adaptive thinking lets Claude decide when to deliberate and how deeply, and the effort parameter is the dial you turn when you want to spend more or less on that deliberation.
A one-pass answer is great for most things and thin for genuinely hard reasoning. Adaptive thinking gives Claude an internal reasoning phase before it replies — it plans, explores, and checks itself, then writes the final answer conditioned on that work. The key word is adaptive: you no longer hand the model a token budget. Claude decides per request whether a problem warrants thinking and how much, and you steer that with a single coarse dial, output_config.effort. The trade-off is unchanged — you pay for every thinking token and wait for it — but the control surface is now one dial per route instead of a number you have to tune per prompt.