Home › Claude Ecosystem Guides
Extended Thinking: How Claude Reasons Before It Answers
Extended thinking is a separate internal reasoning pass a model runs before writing its reply — it works through a hard problem step by step first, then answers, which is different from simply producing a longer response.
What extended thinking actually is
When extended thinking is enabled, the model does not jump straight to an answer. It first generates an internal chain of reasoning — laying out the problem, considering options, checking intermediate steps — and only then commits to a final response. Think of it as scratch work done before the clean answer, not visible clutter added to it.
This matters because many questions are not retrieval problems, they are reasoning problems. A model that answers on its first instinct will often be right on easy tasks and quietly wrong on ones that require several dependent steps. The reasoning pass gives the model room to catch its own mistakes before you ever see the reply.
The raw chain of thought is treated as internal working state and is not exposed as the answer; a readable summary of the reasoning can be surfaced so you understand how the model got there without depending on the exact internal tokens.
Adaptive thinking and effort levels
The modern approach is adaptive: the model itself decides whether a given request needs deep reasoning and how much. A trivial question gets almost no thinking; a multi-constraint design problem gets a lot. You no longer hand-tune a fixed reasoning length for every call — that older "thinking budget" framing has been superseded.
What you control instead is an effort level, from low up to max. Effort is a dial that trades thoroughness against token cost. Low effort keeps things fast and cheap for straightforward work; higher effort tells the model it is allowed to reason harder and longer when the problem earns it.
- Low effort: quick answers, minimal internal reasoning, lowest cost.
- Medium effort: a balance for moderately complex tasks.
- High or max effort: deep, multi-step reasoning for the hardest problems, at higher cost and latency.
When it helps — and when it does not
Extended thinking pays off on tasks with genuine reasoning depth: multi-step math, debugging tangled logic, planning a sequence of tool calls, weighing tradeoffs across constraints, or any question where the first plausible answer is often the wrong one. Because agents chain many decisions together, better upfront planning here compounds — this is where thinking connects to reliable tool use rather than replacing it.
It helps far less on lookups, formatting, short rewrites, or classification, where the answer is essentially known and extra reasoning only burns tokens and time. A concrete example: asked to reconcile a schedule against five overlapping constraints, a non-thinking pass may satisfy three and silently violate two; a thinking pass works the constraints in order and flags the conflict. Same model, different amount of reasoning applied.
The cost and latency tradeoff
Reasoning is not free. The thinking tokens are billed like output tokens, so a high-effort request costs more than a direct answer to the same prompt, and it takes longer before the first visible words appear because the model is working before it writes.
The practical rule is to match effort to difficulty rather than maxing it globally. Reserve high effort for the genuinely hard calls where a wrong answer is expensive, and keep everyday requests on low effort. Because thinking is adaptive, a well-chosen effort level lets the model spend deeply only when a request warrants it, instead of you paying for reasoning on every trivial turn.
How it differs from just longer output
The most common misconception is that extended thinking means "a longer answer." It does not. Length and reasoning are independent: a thinking pass can precede a very short final answer, and a long answer can be produced with no real reasoning behind it at all.
The difference is structural. Extended thinking is a distinct phase — reason first, then respond — whereas telling a model to "be more detailed" only stretches the response phase itself. That is why prompting for verbosity does not reliably improve correctness on hard problems, but a real reasoning pass often does. If you want depth, you are asking for a separate thinking step, not more words in the reply.
Frequently asked
- Is a thinking budget still something I set manually?
- No. The fixed thinking-budget concept has been superseded by adaptive thinking, where the model decides how much to reason, combined with an effort level (low to max) that you set to trade thoroughness against cost.
- Can I read the model's full chain of thought?
- The raw internal reasoning is not exposed as the answer, but a readable summary of the thinking can be shown so you can follow the logic without relying on the exact internal tokens.
- Does extended thinking make every answer better?
- No. It helps on genuinely hard, multi-step reasoning and tool-use planning, but adds cost and latency with little benefit on lookups, formatting, or simple rewrites where the answer is already known.
Work through it interactively
Every area has questions, spaced-repetition cards and a progress record. Those need an account, which is free and takes a moment.
Open the interactive track Create a free account