Claude Extended Thinking

What Is Extended Thinking?

Extended thinking gives Claude a private reasoning scratchpad — a block of tokens it uses to think through a problem before producing its final answer. The reasoning phase is separate from the response: Claude thinks first (you can see this as a thinking content block), then responds.

You control how much thinking budget to allocate via the budget_tokens parameter. Claude uses as many tokens as it needs up to that limit, then writes the final answer. Extended thinking is available on Claude Sonnet 3.7+ and Claude Opus 4.x.

How to Enable Extended Thinking

Python SDK

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 8000  # up to 32,000 max
    },
    messages=[{
        "role": "user",
        "content": "Design the data model and API schema for a multi-tenant SaaS billing system that handles seat-based, usage-based, and hybrid pricing. Include edge cases."
    }]
)

# Response has multiple content blocks
for block in response.content:
    if block.type == "thinking":
        print("=== REASONING ===")
        print(block.thinking)
    elif block.type == "text":
        print("=== RESPONSE ===")
        print(block.text)

TypeScript SDK

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 16000,
  thinking: {
    type: "enabled",
    budget_tokens: 8000,
  },
  messages: [{
    role: "user",
    content: "Find the bug in this recursive algorithm and explain why it fails on edge cases..."
  }]
});

for (const block of response.content) {
  if (block.type === "thinking") {
    console.log("Reasoning:", block.thinking);
  } else if (block.type === "text") {
    console.log("Answer:", block.text);
  }
}

Setting the Right Budget

The budget_tokens parameter is a maximum, not a fixed allocation — Claude uses only what it needs. Setting it too low forces Claude to truncate its reasoning; setting it too high costs money for unused tokens.

When Extended Thinking Helps vs. Doesn't

Cost of Extended Thinking

Thinking tokens are billed as output tokens — the most expensive type. At Claude Sonnet 4.x rates:

For most hard engineering problems, a 4,000–8,000 token budget is the sweet spot: meaningful improvement in reasoning quality without extreme cost. Reserve 32K budgets for genuinely hard math or competitive programming problems.

For current output token prices: prompt-pricing.vercel.app · claude-cost-calc.vercel.app

Reading the Thinking Block

The thinking content block is Claude's internal monologue. It often includes:

• Restating the problem to clarify what's being asked
• Identifying edge cases and failure modes upfront
• Exploring multiple approaches and rejecting weaker ones
• Working through specific examples mentally
• Catching its own mistakes mid-reasoning and correcting

This trace is valuable for debugging: if the final answer is wrong, reading the thinking block often reveals where the reasoning went astray — which you can then correct with a follow-up message.

Extended Thinking in Claude Code (CLI)

When using Claude Code as a CLI tool (not via API), extended thinking is managed automatically. Claude Code decides when to apply extended reasoning based on the complexity of your request — you don't set a budget manually. For the API/SDK use case, the thinking parameter gives you explicit control.

The sub-agents system also benefits from extended thinking: spawning a Plan sub-agent with extended thinking enabled produces more thorough architecture proposals before execution agents begin.

Explore More Claude Code Skills