Extended Thinking

What is extended thinking, and how is it different from how Claude normally answers questions?

Extended thinking is a separate reasoning phase Claude enters before producing its final answer, where it breaks the problem down into smaller pieces, works through them step by step, and checks its own logic for gaps along the way. This reasoning process is typically shown to you, so you can see how Claude got from the question to the answer, rather than just seeing the polished conclusion as usual.

In a normal response, Claude's thinking process stays hidden — you only see the final, organized answer. That makes no difference for simple questions, but for complex problems requiring multi-step reasoning, a hidden thinking process means you can't tell whether Claude missed a key assumption or whether some step in its logic doesn't actually hold up. For example, if you ask Claude to assess the viability of a business decision, extended thinking mode shows you which assumptions it laid out, which options it compared, and why it ruled out a given path — instead of just seeing 'I'd recommend option A.'

What are the limitations or risks of extended thinking, and which one gets misunderstood most often?

The most commonly misunderstood point is: being able to see the reasoning doesn't mean the conclusion is correct. Extended thinking just lays Claude's steps out for you to check, giving you the opportunity to inspect the logic — it doesn't mean that logic is actually sound. If Claude starts from a wrong assumption in the first step, every subsequent step can be perfectly rigorous and the final conclusion will still be wrong. What extended thinking improves is checkability, not a guarantee of correctness — these are two separate things that shouldn't be conflated.

The second commonly overlooked risk is the efficiency trade-off. Extended thinking takes extra processing time, and for questions that are already simple with an obvious answer, forcing it on doesn't make the answer any more accurate — it just slows down the response, which becomes an unnecessary waste of resources. Whether to use extended thinking should hinge on whether the task actually has reasoning worth laying out for inspection, not on the assumption that turning it on is always the safer choice.

When does it make sense to use extended thinking, and when doesn't it?

The core test is whether the answer to this task genuinely needs to be worked out through multiple reasoning steps, rather than looked up directly from existing knowledge. Complex logic puzzles, business decisions that require weighing multiple factors, debugging that requires tracing a problem back to its root cause, or comparing the trade-offs of several options — these are tasks where the value lies in the reasoning process itself, and being able to see that reasoning is what lets you judge whether the conclusion is trustworthy.

It's unnecessary when the answer itself is straightforward — looking up a factual piece of information, a simple format conversion, or routine work where the approach is already clear. For these tasks, turning on extended thinking won't help Claude reason its way to a better answer; it just adds waiting time. A simple test: ask yourself whether you'd need to jot down several steps on paper to work this out yourself, or whether you'd know the answer at a glance. If it needs steps, extended thinking is the right call.

How should advanced users judge whether the reasoning itself in extended thinking actually holds up?

The key move for advanced users is checking the starting point of the reasoning, not just the ending. Every step of the logic can be perfectly rigorous, but if the very first assumption is wrong, no amount of careful reasoning afterward fixes that. In practice, this means first looking at what assumptions or known conditions Claude lays out at the start of its reasoning, confirming whether those assumptions actually hold and whether any important constraint was missed. Only after that check does it become meaningful to evaluate whether the reasoning that follows is logically sound.

A second advanced technique is watching for 'logical jumps' in the reasoning — moving directly from one intermediate conclusion to the next without explaining why that jump is valid. These jumps are especially common in complex tasks, since some steps can look obvious on the surface while actually hiding an untested assumption. When you spot one, you can simply ask Claude to elaborate on that specific step — this is how to actually make use of extended thinking's checkability, rather than accepting the laid-out reasoning at face value.