Extended Thinking

What goes wrong if I set the thinking budget too high?

Two things: time and money. Thinking tokens are billed like output tokens, so a higher budget means potentially more cost — though it is a ceiling, not a guaranteed spend. More immediately, latency: Claude outputs only after finishing thinking, so a higher budget means longer waits. A complex task with a 100K-Token budget can take several times longer than the same question without Extended Thinking.

Size the budget to difficulty: quick questions need no extended thinking; moderate analysis or coding tasks get a few thousand tokens; true multi-step math or long reasoning chains get tens of thousands. Maxing the budget for everything slows fast tasks and costs more unnecessarily.

What observable differences does Extended Thinking actually make?

The most direct effect is more complete structure and layered reasoning. With Extended Thinking on, Claude is more likely to decompose a complex question into sub-problems, reason through each, then integrate — rather than jumping straight to an answer or slipping at a plausible-sounding but logically skipped step.

Another difference shows up on problems with traps. If a question has a non-obvious twist, a model without thinking time is more likely to hit the common error; a model with scratch space is more likely to catch the wrong path during reasoning and self-correct before the final answer. The difference grows with problem complexity and number of steps.

If I can't see the thinking, how do I know Claude is actually reasoning?

You can't directly inspect the thinking, but there are indirect signals. First, the structure and depth of the final answer: if the output has clear step decomposition, handles edge cases, and explains why certain approaches don't work, the thinking was usually doing something useful. Second, some API configurations allow streaming the thinking or showing a summary for spot-checking.

A more practical method: run the same complex question with and without Extended Thinking and compare output quality and error rate. If your task type is consistent, a few comparisons build intuition about how much budget pays off for that type.

Advanced: when does Extended Thinking actually not help?

A few genuine edge cases. First, the task is knowledge retrieval rather than reasoning: if you're asking for a directly look-up-able fact, thinking space doesn't help Claude know things outside its training; it only delays output. Second, the task needs current information: Extended Thinking enhances reasoning, not knowledge recency — if the answer postdates the Knowledge Cutoff, no thinking tokens help.

A subtler third case: very large budgets on certain tasks can cause over-deliberation — the model, having reached the right answer, continues exploring alternatives and occasionally talks itself into a wrong one. For problems with a clear optimal answer, a medium budget tends to be more stable than the maximum.