Claude Extended Thinking: How AI Reasons Through Hard Problems

Bottom line up front: Claude Extended Thinking is a feature that lets Claude show its step-by-step reasoning process before giving a final answer. The model “thinks out loud” — working through a problem carefully — which produces better results on complex math, logic, research, and analysis tasks.

Key Takeaways

• Extended thinking lets Claude reason through hard problems before answering, reducing errors.
• You can see Claude’s reasoning process in a collapsible “thinking” section in the response.
• Most useful for math, logic, multi-step analysis, coding problems, and strategic planning.
• Available in Claude.ai on Pro and Max plans, and via the Anthropic API.
• Works best with Claude Opus 4.6 but is also available with Claude Sonnet 4.6.

What Is Extended Thinking?

Extended thinking is a mode Claude can enter where it works through a problem step by step before committing to a final answer. In a standard response, Claude generates an answer directly. With extended thinking turned on, Claude first produces a detailed chain of reasoning — exploring the problem, considering different angles, catching potential errors — and then gives you a final answer informed by all of that work.

You see this reasoning in the interface. There’s a collapsible “Thinking” block above Claude’s response. You can expand it to see exactly how Claude approached the problem. This transparency is valuable not just for the improved accuracy — it also lets you check Claude’s logic, spot where it might have gone wrong, and learn from how it breaks down complex problems.

The concept is related to what researchers call “chain-of-thought prompting” — the idea that getting an AI to show its work produces better results than having it jump straight to an answer. Anthropic has built this into Claude as a first-class feature rather than something you have to engineer through clever prompting.

Why Does Extended Thinking Improve Results?

The improvement comes from giving the model more “compute time” to work on the problem. A standard answer is generated left to right, one token at a time, with no opportunity to backtrack or reconsider. Extended thinking creates a separate reasoning space where Claude can explore, self-correct, and build up to an answer systematically.

Think about how you solve a hard math problem. You don’t just write down the answer immediately — you work through it, check your arithmetic, notice if something looks wrong, and revise. Extended thinking gives Claude a version of that same process. The result is fewer errors, more complete answers, and better performance on tasks that require multiple steps of logic.

Anthropic has published benchmark results showing meaningful gains from extended thinking on tasks like math competition problems (AIME), graduate-level science questions (GPQA), and complex coding challenges. The gains are largest for the hardest problems — the ones where getting every step right matters most.

When Should You Use Extended Thinking?

Extended thinking isn’t always necessary. It adds latency — Claude takes longer to respond when it’s thinking through a problem in depth. For simple tasks, that tradeoff isn’t worth it. Here’s a practical guide:

Turn on extended thinking for:

• Multi-step math problems where every step needs to be correct.
• Logic puzzles, probability problems, and deductive reasoning tasks.
• Complex code reviews where you need Claude to check for subtle bugs or architectural issues.
• Strategic planning tasks where you need a thorough exploration of options and tradeoffs.
• Research synthesis where you need Claude to weigh conflicting sources carefully.
• Legal and financial analysis where precision matters and errors have real consequences.

Skip extended thinking for:

• Simple Q&A and factual lookups.
• Everyday writing and editing tasks.
• Casual conversation and brainstorming.
• Tasks where you need a fast response and the problem is not complex.

How to Turn On Extended Thinking

In Claude.ai

Extended thinking can be toggled on or off in the Claude.ai interface. Look for the “Extended thinking” or “Think deeper” toggle in the conversation settings. When enabled, Claude’s next response will include a reasoning section. You can see it by expanding the “Thinking” block at the top of the response.

Via the Anthropic API

Developers enable extended thinking by passing a thinking configuration in the API request. You set a budget_tokens parameter that controls how much compute Claude uses for reasoning — more tokens means deeper thinking but higher cost and latency. Here’s a simplified example of how it looks in the API:

response = client.messages.create(
    model="claude-opus-4-6-20250514",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000
    },
    messages=[{"role": "user", "content": "Solve this step by step: ..."}]
)

The thinking blocks appear in the response alongside the text blocks. You can display them to users or process them internally depending on your use case. For more on building with Claude’s API, see our Claude Agent SDK guide.

Extended Thinking With Opus vs. Sonnet

Extended thinking is available on both Claude Opus 4.6 and Claude Sonnet 4.6, but the performance gains are largest with Opus. Opus has deeper reasoning capabilities to begin with, so extended thinking amplifies those strengths on the hardest problems. Sonnet with extended thinking is still a significant upgrade over standard Sonnet for moderately complex tasks.

Read our dedicated guides for both models: Claude Opus 4.6 explained and Claude Sonnet 4.6 explained. For a plan comparison that includes extended thinking access, see our Claude plans guide.

Real-World Examples

Complex Math and Statistics

A data analyst uses extended thinking to work through a complex statistical analysis. Standard Claude might make an arithmetic error or skip a step in the calculation. Extended thinking Claude shows every step of the work, making it easy to verify the answer and catch any issues.

Legal Document Review

A lawyer asks Claude to identify every clause in a contract that could create liability. With extended thinking, Claude methodically goes through each section, flags specific language, and explains its reasoning for each flag. The result is much more thorough than a standard scan. See how lawyers are using AI tools in their practice.

Software Architecture

A developer asks Claude to review a proposed system architecture and identify weaknesses. Extended thinking lets Claude consider scalability, security, failure modes, and maintenance burden in a structured way before giving recommendations. Developers working with Claude Code benefit from this — read our Claude Code guide for more.

Strategic Business Planning

A consultant uses extended thinking to analyze a complex market entry decision. Claude considers competitive dynamics, regulatory factors, customer segments, and financial modeling systematically before giving a recommendation. For consulting use cases, see our AI for consultants guide.

Limitations of Extended Thinking

Extended thinking adds latency. For complex problems, the thinking phase can take several seconds to a minute. This is a real cost for real-time applications or impatient users. On the API, extended thinking also uses more tokens (the thinking tokens are counted), which increases costs.

Extended thinking also doesn’t guarantee perfect answers. It reduces errors on hard problems, but Claude can still make mistakes. The visible reasoning helps you catch those mistakes, which is part of the value — but you should still review outputs on critical tasks.

How It Fits Into Your Claude Workflow

For new users, start with the Claude beginner’s guide to understand the basics before diving into extended thinking. If you use Claude via the desktop app, see our Claude desktop app guide to learn how to access extended thinking there. Extended thinking is also relevant for developers building AI agents — see our Agent SDK guide for how to incorporate it into automated workflows.

External Resources

For academic background on chain-of-thought reasoning in large language models, see the Wikipedia article on chain-of-thought prompting. Anthropic’s research blog post on extended thinking has technical details. For the foundational paper that inspired this approach, see the original chain-of-thought prompting paper on arXiv.

Free Resource: Extended Thinking Prompt Templates

We’ve put together a free pack of 20 Extended Thinking prompt templates for math, logic, research, and analysis tasks. Download it free in our products library and get Claude reasoning at its best immediately. Get the Extended Thinking Prompt Pack →

Stay Updated

Anthropic continues to expand extended thinking capabilities with each model update. New integrations, improved reasoning depth, and lower latency are all in development. Join the Beginners in AI newsletter for the latest updates in plain English. Subscribe here →

Frequently Asked Questions

Does extended thinking always give better answers?

Not always. For simple tasks, standard Claude is fine. Extended thinking shines on complex, multi-step problems where intermediate reasoning steps matter. Using it on a simple task adds latency without meaningful quality gain.

Can I turn off extended thinking mid-conversation?

Yes. In Claude.ai, you can toggle extended thinking on or off for each message. On the API, you specify it per request. You’re not committed to using it for a whole conversation.

Is the thinking content private or used for training?

Anthropic’s standard data handling policies apply. On Claude.ai, your conversations are not used to train models by default if you opt out in settings. The thinking content is treated the same as the rest of your conversation.

Does extended thinking cost more on the API?

Yes. The thinking tokens are counted as output tokens and billed accordingly. On Claude Max (subscription), you don’t pay per token, so extended thinking has no extra cost beyond the time it takes.

Can extended thinking help with creative writing?

Somewhat. Extended thinking is most valuable for analytical and logical tasks. For creative writing, the benefit is smaller, though it can help Claude plan a complex story structure or develop a nuanced argument before writing. Most writers don’t need to turn it on for standard creative tasks.

You May Also Like

• How to Use Claude AI
• Best Claude Prompts
• Claude AI Login
• ChatGPT vs Claude vs Gemini
• Claude Code Beginners Guide

---

Get free AI tips daily → Subscribe to Beginners in AI

Sources

This article draws on official documentation, product pages, and industry reporting. Specific sources are linked inline throughout the text.

Last reviewed: April 2026