Claude’s extended thinking mode allows the AI to work through complex problems step by step before giving you a final answer.
Like a mathematician who shows their work, Claude in extended thinking mode reveals its reasoning process before presenting a conclusion.
This guide explains what extended thinking is, how it works under the hood, when to enable it, and how to get the most from it.
What Is Claude’s Extended Thinking Mode and How It Works
Claude’s extended thinking mode is a capability that allows Claude to spend more computation time reasoning through a problem before responding.
In standard mode, Claude produces a response in a single pass. In extended thinking mode, it first generates a chain of reasoning tokens.
These thinking tokens are internal deliberation steps where Claude breaks down the problem, explores different approaches, and checks its work.
After thinking through the problem, Claude synthesizes its reasoning into the final response you see, which is often dramatically more accurate.
Extended thinking is the same Claude model given more compute time, not a separate model. The quality improvement comes from deeper deliberation.
When you enable thinking in Claude, a collapsible reasoning section appears above the final answer showing Claude’s step-by-step thought process.
This visible reasoning is one of the most distinctive features. You can read how Claude approached the problem, which builds confidence in the answer.
The thinking content shows how Claude broke down the problem into parts, identified the relevant information, and arrived at its conclusion.
Claude Opus 4.8 and Claude Sonnet 4.6 use adaptive thinking, where Claude dynamically decides how much to think based on question complexity.
Adaptive thinking means simple questions get fast, lean responses while hard questions automatically trigger deeper reasoning without you adjusting settings.
Claude Fable 5, Anthropic’s most capable model, always thinks. Its thinking cannot be disabled, and the visible summary is provided automatically.
According to Anthropic API documentation, thinking tokens are billed at the same rate as output tokens for the model being used.
This means enabling extended thinking increases cost per query, which is an important consideration for high-volume API applications.
For claude.ai users on paid plans, thinking mode is available through the interface controls without direct token cost management concerns.
When to Enable Extended Thinking in Claude for Best Results
Extended thinking delivers the most value on tasks where careful reasoning, logical consistency, and step-by-step problem decomposition matter most.
Mathematics is one of the strongest use cases: complex proofs, multi-step calculations, and algebraic problems all benefit from extended thinking.
Logic puzzles, constraint satisfaction problems, and riddles that require systematic exploration of possibilities also show dramatic improvement.
Complex coding tasks benefit too: debugging subtle logic errors, designing data structures, and reasoning about algorithmic complexity improve with thinking.
Legal and contractual analysis is excellent for extended thinking, especially multi-party contract review where conflicting clauses need resolution.
Strategic planning tasks where Claude must weigh multiple competing factors and trade-offs produce noticeably better analysis with extended thinking.
Scientific reasoning tasks, like explaining a mechanism with multiple interacting variables, also benefit from Claude’s structured deliberation.
Extended thinking is NOT necessary for simple questions, general writing assistance, basic information requests, or creative brainstorming.
Using it for simple tasks wastes token budget and adds unnecessary latency without any meaningful improvement in response quality.
A good rule of thumb: if a human expert would think for several minutes before answering, extended thinking is likely worth enabling for that task.
Conversely, if a human expert would answer in seconds from memory, standard Claude mode is sufficient and much faster to use.
Claude’s adaptive thinking system handles many of these decisions automatically, but you can always override and force extended thinking on or off.
See our Claude prompting guide for tips on structuring questions to maximize the quality of Claude’s extended thinking output.
The tasks where extended thinking pays off most are complex, high-stakes questions where a wrong answer has significant consequences for you.
How to Turn On Extended Thinking in the Claude Interface
Enabling extended thinking in claude.ai is straightforward. The control is available in the interface for users on paid plans with access to capable models.
In the Claude web app, look for the ‘Thinking’ toggle in the chat toolbar or the model selector, depending on your interface version.
Clicking the Thinking toggle activates it for your current conversation. Claude will show its reasoning in an expandable section for each response.
The effort level slider, available alongside the thinking toggle, controls how thoroughly Claude reasons. Higher effort means longer, deeper thinking.
Setting effort to ‘low’ uses thinking sparingly, giving a light boost on moderate questions without incurring the full latency of deep reasoning.
Setting effort to ‘high’ or ‘max’ triggers the most thorough reasoning, best for genuinely hard problems where accuracy matters most.
For everyday complex questions, a medium effort level usually provides the best balance between response quality and speed for most users.
In the Claude desktop app, the thinking control is similarly accessible from the chat input area or the response settings panel interface.
Claude mobile apps support extended thinking on current models, with the thinking toggle visible in the chat options when available for your plan.
The thinking section in responses is collapsible. You can read it to understand Claude’s reasoning or collapse it to see just the final answer.
Even unread, the thinking section still improves the final answer. The reasoning enhances the conclusion whether you view it or not.
For Claude Code, the terminal agent, you can enable extended thinking using the ‘think’ command: type ‘claude –think’ in your terminal session.
Enterprise users can set default thinking levels across their organization’s Claude deployment. See how Claude Pro and Max plans differ for thinking access.
Once you experience extended thinking on a genuinely hard problem, you will understand immediately why it produces qualitatively better reasoning.
What Extended Thinking Looks Like in a Real Claude Conversation
Seeing a real example of Claude extended thinking in action makes the feature much easier to understand and appreciate in concrete terms.
Take this example: ‘A train leaves New York at 3pm going 90mph west. Another leaves Chicago at 5pm going 75mph east. When do they meet?’
In standard mode, Claude might jump to a calculation and occasionally make arithmetic errors on multi-step problems like this one.
In extended thinking mode, Claude first writes out its understanding of the problem, identifies the variables, sets up the equation, and solves step by step.
The thinking section might show: ‘Distance NY to Chicago is approximately 790 miles. Train 1 travels for t hours. Train 2 travels for (t-2) hours…’
Claude continues showing its algebra, checks the result by substituting back, confirms it is correct, and only then writes the final clean answer.
On complex coding tasks, the thinking section shows Claude reading through the bug description, identifying candidate causes, and reasoning through each one.
This visible elimination process lets you follow along and correct Claude if it misidentifies the root cause before it writes the fix.
For strategic analysis, the thinking section reveals how Claude weighs competing factors, considers counterarguments, and arrives at its recommendation.
Seeing this reasoning makes it much easier to spot if Claude made incorrect assumptions that you can quickly correct for a better final answer.
The visibility of reasoning is a major trust-building feature. You can evaluate Claude’s logic rather than just accepting its conclusion on faith.
Researchers and analysts particularly value this transparency for auditing Claude’s reasoning on high-stakes analytical or technical questions.
Our context window guide explains why large context capacity is important for extended thinking on long, complex tasks.
The combination of transparent reasoning and high accuracy makes extended thinking Claude’s most powerful mode for serious professional work.
Extended Thinking vs Standard Claude Responses: A Comparison
Comparing extended thinking and standard Claude responses reveals when each mode is clearly superior and where the difference is negligible.
For creative writing, standard Claude is usually better. Extended thinking can overthink creative tasks and produce more formulaic output.
For math, logic, and technical problem-solving, extended thinking consistently produces more accurate results with fewer silly mistakes.
For general knowledge questions with factual answers, standard Claude is fast and accurate enough. Extended thinking adds cost without benefit.
For complex multi-variable analysis, extended thinking produces noticeably more nuanced and balanced conclusions than standard mode typically delivers.
Response length in extended thinking mode tends to be more concise in the final answer, since Claude has resolved uncertainties during thinking.
Standard mode sometimes hedges more because Claude is uncertain but unable to reason through the uncertainty before committing to output.
Extended thinking reduces visible hedging because Claude has already worked through the ambiguity internally before writing the final response.
Latency is the primary cost of extended thinking. Responses take longer to arrive because Claude is spending extra time in the reasoning phase.
For interactive applications where fast responses matter, standard mode is preferable except for specific tasks that genuinely need deep reasoning.
For batch processing, analysis pipelines, and offline tasks where latency is not critical, extended thinking is almost always worth enabling.
Token costs are higher with extended thinking because thinking tokens are billed alongside the output tokens for the same model price per token.
Adding 5 to 15 seconds of latency and a fraction of a cent per call, the trade-off is almost always worth it for complex tasks.
The right approach: use standard mode by default and enable extended thinking only when the task genuinely demands deeper reasoning.
Using Claude Extended Thinking via the API for Developers
Developers can integrate Claude’s extended thinking capability into their applications via the API using the thinking parameter on messages.
On Claude Opus 4.8 and Sonnet 4.6, include ‘thinking: {type: adaptive}’ in your API request to enable adaptive thinking for that call.
Adaptive thinking lets Claude decide how much to think based on the question, which is the recommended setting for most production applications.
You can also set a specific budget_tokens value to cap the maximum tokens Claude spends on reasoning, controlling both cost and latency tightly.
API responses with thinking enabled return thinking content blocks in the content array, followed by the regular text block with the final answer.
Parse the thinking content type to extract and display Claude’s reasoning process in your application’s UI if that transparency serves your users.
For Claude Fable 5, thinking is always on and cannot be disabled. Use the effort parameter to control reasoning depth from low to max levels.
Avoid passing modified thinking blocks back to Claude in subsequent messages. Return them exactly as received or omit them, but never edit the content.
Thinking blocks from one model should not be passed to a different model. They are model-specific and will be silently dropped if passed to another.
For streaming API calls, thinking tokens stream before the text response. Design your streaming UI to handle the thinking block type gracefully.
Review the extended thinking prompting tips at Anthropic docs for production-ready guidance on structuring prompts for thinking mode.
Test your application’s UX for the additional latency that thinking adds. Implement loading indicators or progress messages to keep users informed.
Extended thinking is one of the most powerful tools in the Claude API for applications where answer accuracy is more critical than response speed.
The combination of visible reasoning and improved accuracy makes it the mode of choice for any application solving complex, high-stakes problems.
Related Articles
Trust Post Desk
A journalist and editor at TrustPost.org covering world and national news, technology updates and human-interest stories. They check every fact, interview sources in person or online, and aim to deliver clear, accurate reporting. Their work ranges from breaking news to in-depth features and daily newsletters. Outside the newsroom, they follow emerging trends and engage with readers on social media.
