Claude Extended Thinking Mode Tutorial: Step-by-Step Guide (2026)

Most people using Claude are leaving its most powerful feature completely untouched. Extended Thinking mode lets Claude spend significantly more time reasoning through complex problems before giving you an answer — and the difference in output quality is dramatic. I’ve run it on everything from multi-step business analyses to debugging tangled code logic, and the results consistently beat standard Claude responses on hard problems by a wide margin.

A BCG study found that knowledge workers using AI completed tasks 25–40% faster on average.

This tutorial walks you through exactly how to turn it on, configure it correctly, and use it effectively — with real prompt templates and specific settings you can copy right now.

What You’ll Build by the End of This Tutorial

By the time you finish reading and following along, you’ll be able to:

• Enable and configure Extended Thinking mode inside Claude.ai
• Access Extended Thinking via the Anthropic API with the correct parameters
• Write prompts that actually take advantage of the extended reasoning window
• Know when to use it (and when it’s overkill)
• Avoid the most common setup mistakes that waste tokens and time

Prerequisites: What You Need Before Starting

Before we get into the steps, make sure you have the following:

• A Claude Pro or Claude Max account — Extended Thinking is not available on the free tier. Claude Pro runs $20/month. Claude Max starts at $100/month and gives you significantly higher usage limits.
• Access to Claude.ai — or an Anthropic API key if you want to integrate Extended Thinking into your own tools or workflows.
• A real problem to solve — Extended Thinking shines on complex, multi-part questions. Have one ready so you can test it live as you follow along.
• Basic familiarity with Claude — you don’t need to be a power user, but you should know how to start a conversation and write a prompt.

If you’re using the API route, you’ll also need a basic understanding of JSON and HTTP requests, or a tool like Postman or the Anthropic Python SDK.

Step 1 — Log In and Open a New Claude Conversation

Go to claude.ai and log in with your Pro or Max account. Click “New Chat” in the left sidebar to open a fresh conversation window.

Make sure you’re on Claude Sonnet or Claude Opus — Extended Thinking is tied to these models. At the time of writing in 2026, Claude Sonnet 4 and Claude Opus 4 both support Extended Thinking. You’ll see the model selector at the top of the chat window. Click it and confirm you’re not on a legacy or lite model.

Step 2 — Enabling Extended Thinking Mode in Claude.ai

Look at the bottom of the chat input box. You’ll see a row of small icons — one of them looks like a small brain or a lightbulb, depending on your interface version. This is the Extended Thinking toggle.

Click it. A small tooltip or dropdown will confirm that Extended Thinking is now active for this conversation. The icon will highlight or change color to indicate it’s on.

What changes when you enable it:

• Claude will show you a “Thinking…” section above its final answer — this is the internal reasoning chain made visible
• Response time increases, sometimes by 30–90 seconds on complex prompts
• The answers are noticeably more structured and self-correcting
• Token usage goes up, which counts against your plan limits

I tested this back-to-back on a 12-step business valuation problem. Standard Claude gave me a decent answer in 8 seconds. Extended Thinking took 47 seconds — and caught two logical errors in its own reasoning before presenting the final output. That self-correction alone made it worth it.

Step 3 — Understanding the Thinking Budget (Critical Setting)

This is the setting most tutorials skip, and it’s actually the most important one if you’re using the API.

Extended Thinking uses a parameter called thinking.budget_tokens. This controls how many tokens Claude can use for its internal reasoning before it writes the visible response.

Budget Setting	Token Range	Best For	Response Time
Low	1,000–5,000	Simple analysis, short summaries	10–25 seconds
Medium	5,000–10,000	Research synthesis, code review	25–60 seconds
High	10,000–32,000	Complex strategy, multi-step reasoning	60–180 seconds
Maximum	32,000+	Frontier reasoning tasks, research	2–5+ minutes

In the Claude.ai interface, Anthropic manages this automatically based on your subscription. In the API, you set it explicitly. Don’t just crank it to maximum — higher budgets don’t always mean better answers. Match the budget to the actual complexity of your task.

Step 4 — Setting Up Extended Thinking via the Anthropic API

If you want to build Extended Thinking into your own workflows, automations, or apps, here’s the exact API setup. This uses the Anthropic Python SDK.

First, install the SDK if you haven’t already:

pip install anthropic

Then, here’s a working code snippet with Extended Thinking enabled:

import anthropic

client = anthropic.Anthropic(api_key="your-api-key-here")

response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000
    },
    messages=[
        {
            "role": "user",
            "content": "Your complex prompt goes here"
        }
    ]
)

# Access the thinking block
for block in response.content:
    if block.type == "thinking":
        print("THINKING PROCESS:")
        print(block.thinking)
    elif block.type == "text":
        print("FINAL ANSWER:")
        print(block.text)

Critical notes on these settings:

• max_tokens must be greater than budget_tokens — Claude needs token space for both the thinking and the final response
• The thinking block in the response is separate from the text block — parse them individually
• Extended Thinking is not compatible with streaming in all configurations — check Anthropic’s current docs if you’re building a real-time interface
• Temperature is automatically set to 1 when Extended Thinking is active — you can’t change it

Step 5 — Writing Prompts That Actually Benefit From Extended Thinking

This is where most people go wrong. They turn on Extended Thinking and then ask it something like “write me a blog post intro.” Extended Thinking doesn’t help with that. It’s built for tasks where reasoning depth matters.

Here are the prompt structures that consistently get the best results:

Prompt Template 1: Strategic Analysis

You are a strategic business advisor. I need a thorough analysis of the following situation.

CONTEXT:
[Describe your situation in specific detail — industry, size, current state]

CONSTRAINTS:
[List your actual limitations — budget, team size, timeline, technical skills]

QUESTION:
[Ask one focused, high-stakes question]

Please reason through this carefully before answering. Identify assumptions you're making, consider at least 3 alternative approaches, and flag any risks in your recommended path.

Prompt Template 2: Complex Code Debugging

I have a bug in the following code and I can't figure out why it's happening. 

LANGUAGE: [Python / JavaScript / etc.]

CODE:
[Paste your full code block here]

EXPECTED BEHAVIOR:
[What should it do]

ACTUAL BEHAVIOR:
[What it actually does, including any error messages]

WHAT I'VE ALREADY TRIED:
[List 2-3 things you tested]

Please trace through the logic step by step, identify the root cause (not just symptoms), and give me a fix with an explanation of why the original code failed.

Prompt Template 3: Research Synthesis

I need you to synthesize the following information into a coherent, prioritized set of recommendations.

SOURCES / DATA:
[Paste your raw notes, data points, or multiple documents here]

MY GOAL:
[One clear sentence about what you're trying to decide or achieve]

AUDIENCE FOR THE FINAL RECOMMENDATION:
[Who will read this — yourself, a client, a board, etc.]

Please identify patterns across the sources, note any contradictions, weigh the evidence, and produce a ranked list of actionable recommendations with your reasoning for each ranking.

In my experience, adding the phrase “Please reason through this carefully before answering” consistently produces more thorough thinking blocks, even when Extended Thinking is already enabled. It’s redundant in theory, but it seems to signal to the model that depth is expected.

Step 6 — Reading the Thinking Block (What to Actually Look For)

When Extended Thinking is on, Claude shows you its reasoning chain in a collapsible “Thinking” section above the answer. Most people glance at it and move on. That’s a mistake — the thinking block is genuinely useful.

Here’s what to look for when you read it:

• Self-corrections — if you see Claude write something and then say “wait, that’s not right” or “actually, reconsidering this,” that’s the feature working. It caught an error before it reached you.
• Assumptions it’s making — Claude will often surface assumptions in the thinking block that it doesn’t explicitly state in the final answer. These are worth reviewing, because they might be wrong for your situation.
• Paths it rejected — sometimes the most valuable insight is why Claude ruled out an approach. That reasoning might reveal something useful you hadn’t considered.
• Where it got uncertain — if the thinking block shows the model going back and forth on something, that’s a signal the question is genuinely ambiguous and you should add more context.

Step 7 — When to Use Extended Thinking (and When to Skip It)

Extended Thinking is not a universal upgrade. Using it for simple tasks wastes your token budget and adds unnecessary wait time. Here’s a clear breakdown:

Task Type	Use Extended Thinking?	Why
Business strategy / decision analysis	✅ Yes	Multiple variables, high stakes, needs trade-off reasoning
Complex code debugging	✅ Yes	Logic tracing benefits from extended reasoning chains
Math / quantitative problem solving	✅ Yes	Step-by-step verification catches arithmetic errors
Research synthesis across sources	✅ Yes	Contradiction detection and pattern finding improve
Writing a blog post / email	❌ No	Creative output doesn’t require deep logical reasoning
Summarizing a document	❌ No	Standard Claude handles this fine — overkill here
Generating social media captions	❌ No	Minimal benefit, wastes your monthly token budget
Legal / contract analysis	✅ Yes	Clause interactions require careful multi-step reasoning

Common Troubleshooting: When Extended Thinking Isn’t Working Right

Problem: The thinking block appears but the answer is still shallow

Fix: Your prompt is probably too vague. Extended Thinking needs something to actually reason through. Rewrite your prompt to include specific context, constraints, and a focused question. Vague input produces vague reasoning, even with more thinking time.

Problem: Getting an API error when enabling Extended Thinking

Fix: Check that your max_tokens value is higher than your budget_tokens value. This is the most common API mistake. If budget_tokens is 10,000 and max_tokens is 8,000, the request will fail. Set max_tokens to at least budget_tokens + 4,000 to give the final response enough room.

Problem: Extended Thinking toggle isn’t visible in the Claude.ai interface

Fix: Confirm you’re on a Pro or Max subscription and that you’re using a compatible model (Sonnet or Opus series). If you just upgraded your plan, try logging out and back in. The toggle sometimes doesn’t appear on older chat threads — start a fresh conversation.

Problem: Response takes too long and times out

Fix: Lower your thinking budget. If you’re using the API and hitting timeouts, you may also need to implement streaming or increase your HTTP timeout settings. For the claude.ai interface, very high budget requests on slow connections can stall — try refreshing and starting a new thread with a lower-complexity version of the prompt first.

Problem: Claude says it can’t use Extended Thinking for this request

Fix: A small number of task types trigger safety or policy guardrails that disable extended reasoning. This is rare. If it happens, rephrase the prompt to be more specific about the professional or analytical nature of the task.

“`html

My Real-World Experience

A few months ago I had a situation that made me actually stop and think about what I was doing. A couple from Germany came to me wanting to buy a 3-bedroom apartment in Funchal — their budget was tight, the market was moving fast, and they needed a solid comparative market analysis before making an offer. Normally that report takes me the better part of an afternoon: pulling recent sales data, writing up the neighbourhood context, justifying the price range, all of it. I decided to run the whole thing through Claude with Extended Thinking turned on. I gave it my raw data — recent comparable sales, property specs, location notes — and asked it to reason through the valuation and structure the full CMA. What came back wasn’t just faster. It actually caught a pricing inconsistency I’d missed between two comparable properties with different parking situations. The final report took me 40 minutes instead of my usual 3 hours. That one session alone justified the monthly cost.

I’ve since used Extended Thinking mode for about 6 weeks across different tasks — drafting complex client emails where I need to say something difficult without losing the lead, building out neighbourhood reports for Câmara de Lobos and Caniço, and working through ad targeting logic for my Facebook campaigns. The reasoning it does before answering is genuinely different from the standard mode. You can feel it working through the problem rather than just pattern-matching a response.

That said, one real frustration: the response time. When Extended Thinking is active on a complex prompt, you’re sometimes waiting 45 to 90 seconds for the output. When I’m in the middle of a busy morning — phone ringing, WhatsApp notifications stacking up — that pause feels long. It’s not a dealbreaker, but it does mean I schedule this work for focused blocks rather than quick in-between tasks.

Rating for solo real estate use: 4.5/5 — the depth of reasoning it brings to property analysis and client communication is something I genuinely haven’t found in any other AI tool I’ve tested at this price point.

Bottom line: If you’re a solo agent drowning in reports, client follow-ups, and market research like I was, Extended Thinking mode will save you real hours every week. I’d recommend it without hesitation to any one-person real estate operation that can’t afford to hire an analyst.

“`

Real-World Example: Solving a Pricing Strategy Problem With Extended Thinking