Claude's "Thinking Mode" Is the Most Underused Feature in AI Right Now
Most people open Claude, type a question, and get an answer. That's fine — but they're skipping a feature that fundamentally changes what Claude is capable of. Extended Thinking is Claude's built-in reasoning mode, and when you activate it, you're not just getting a longer response — you're getting a different cognitive process altogether. By the end of this article, you'll know exactly what it is, why it produces better outputs, and how to use it today on problems that actually matter to you.
Extended Thinking Isn't a Setting — It's a Completely Different Mode of Reasoning
Most AI tools, including standard Claude, work in a straight line: you send a message, the model predicts the next best token, and you get an answer. It's fast and often impressive. But it's also shallow in ways that matter when the problem is genuinely complex.
Extended Thinking tells Claude to reason through a problem before it gives you an answer. It's similar to how you'd solve a hard math problem by writing out the steps on scratch paper instead of trying to do it all in your head. Claude generates an internal chain of reasoning — a "scratchpad" — that it uses to check assumptions, explore alternatives, and self-correct before the final response lands in your chat.
Anthropic built this feature specifically for tasks where fast pattern-matching fails: multi-step logic problems, ambiguous decisions, technical analysis, and anything requiring the model to hold multiple competing ideas at once. This isn't a marketing claim — in Anthropic's own benchmarks, extended thinking pushed Claude's performance on hard math and science problems significantly above what standard mode could produce.
The practical difference is real. Ask standard Claude "What's the flaw in this business model?" and you'll get a polished, surface-level critique. Ask Extended Thinking Claude the same question, and it will find contradictions you missed, stress-test assumptions, and often give you a conclusion it initially disagreed with — because it reasoned its way to a better answer.
The "Scratchpad" Is the Secret — And Most People Never See It
Here's what makes Extended Thinking genuinely different from just asking Claude to "think step by step." That classic prompt trick — popularized in a 2022 Google research paper — does improve outputs, but it's still Claude performing for you. It's showing its work after the fact, not actually changing how it processes your request.
Extended Thinking happens before Claude writes its response. The model generates what Anthropic calls a thinking block — a messy, exploratory internal monologue where Claude can be wrong, change its mind, and backtrack. You can actually see this in Claude.ai when the feature is active. It shows up as a collapsible section labeled "Thinking" that appears above the final answer.
Reading that thinking block is one of the most valuable things you can do. It tells you exactly where the model is uncertain, what trade-offs it considered, and what it ruled out. That's information you can use — to ask better follow-up questions, to spot where your original prompt was ambiguous, or to identify assumptions worth challenging.
Think of it like getting the analyst's working notes alongside their final report. The report might say "Option B is better." The notes show you it was genuinely close, that Option A wins under one specific scenario, and that the whole conclusion depends on a variable you haven't confirmed yet. That context changes everything about how you act on the answer.
A smart way to use this: after Extended Thinking gives you a response, read the thinking block first, then ask "Based on your reasoning, what would need to be true for the opposite conclusion to be correct?" — Claude will give you a devastatingly useful answer because the groundwork is already laid.
How to Actually Activate Extended Thinking and Use It Well
Extended Thinking is available on Claude.ai for Pro subscribers, and through the Anthropic API for developers. On the API, you enable it by setting "thinking": {"type": "enabled", "budget_tokens": [number]} in your request — the token budget controls how long Claude can think before it responds.
On Claude.ai, the process is simpler. Open a new conversation, click the icon that looks like a brain or a lightbulb (it varies slightly by interface update), and toggle Extended Thinking on. You'll know it's working when you see the "Thinking" block appear before Claude's response.
Here's where to start — a three-step workflow you can run in the next 15 minutes:
Step 1: Bring a real problem. Don't test it with a trivia question. Bring something that genuinely has layers — a career decision, a technical architecture choice, a strategic tradeoff. Try: "I'm deciding between [Option A] and [Option B] for [specific situation]. Walk me through the key tradeoffs, flag any assumptions you're making, and tell me what information would most change your recommendation."
Step 2: Read the thinking block before the answer. Seriously, start there. Look for the moments Claude changes its mind or says something like "wait, that's not right." Those inflection points are where the real value lives.
Step 3: Use the uncertainty as your next prompt. Wherever Claude expressed doubt in the thinking block, push on it. Ask: "You seemed uncertain about [X]. What would it take to resolve that uncertainty?" This turns one response into a genuine back-and-forth reasoning session.
Extended Thinking also pairs exceptionally well with long documents. Paste in a contract, a business plan, or a research paper and ask Claude to identify the three riskiest assumptions the author is making. Standard Claude will give you something competent. Extended Thinking Claude will actually read against the grain.
The Part Most People Get Wrong
Most people treat Extended Thinking like a "better answers" button they click for everything. That's wrong, and it'll frustrate you fast.
Extended Thinking is slower and uses significantly more compute. On the API, thinking tokens cost real money. On Claude.ai, the response takes noticeably longer. If you're using it to write a quick email or summarize a short article, you're wasting the feature — and probably annoyed at the wait time.
The right mental model: use Extended Thinking when the cost of a wrong answer is high. Before a big client presentation. When you're debugging logic in a complex system. When you're making a decision you'll live with for months. That's when the extra reasoning cycles pay off.
The other common mistake is ignoring the thinking block entirely and just reading the final answer. People do this because the final answer looks clean and the thinking block looks messy. But "messy" is exactly the point — that's where Claude is being honest with itself. Skipping it means you're back to treating Claude like a search engine with better prose.
Key Takeaways
- Extended Thinking: A reasoning mode where Claude generates an internal "scratchpad" before responding, producing more accurate and nuanced outputs on complex problems.
- Thinking Block: The visible internal monologue you can read in Claude.ai — it shows uncertainty, self-correction, and trade-offs that the final answer may not surface.
- Best Use Cases: High-stakes decisions, multi-step logic, technical analysis, and document critique — not everyday quick tasks.
- The Follow-Up Prompt: After reading the thinking block, always ask what would need to be true for the opposite conclusion — this unlocks Claude's most useful reasoning.
- Access: Available on Claude.ai Pro and the Anthropic API; enable it intentionally and match the tool to the task, not the other way around.
What to Do Right Now
Open Claude.ai, switch on Extended Thinking, and paste in one real decision you're sitting on — a business choice, a project direction, a technical problem. Start with this prompt: "I'm trying to decide [your situation]. Think through this carefully, tell me what you're uncertain about, and flag what additional information would change your answer most." Read the thinking block first, then the answer, and I promise you'll use it differently from this point forward.