Extended thinking

Extended thinking is the feature that lets Claude reason before responding. The response contains thinking blocks ahead of the answer text, and those thinking tokens count toward output usage and are billed at the output rate. Other providers have their own versions of the same idea under different names.

On current Claude models, thinking is adaptive by default: the model decides per request whether and how much to think, and you steer it with the effort setting rather than a fixed token budget, which current models reject. By default the thinking text itself is omitted (the blocks arrive empty), and you can request a readable summary if you want to show progress or debug. Some models let you turn thinking off for latency-sensitive work; Opus 5.5 does not, so you lower effort instead.

In a product, the practical concerns are latency, cost, and length. Users see nothing until thinking ends unless you stream status, so show a working indicator or a summarized thought. Track output tokens, since thinking can be a large share of them. And set max_tokens generously, because thinking and answer share that ceiling and a tight cap can cut the answer off.