markus41/plugins/claude-code-expert/skills/extended-thinking

Claude Code Extended Thinking

Complete guide to extended thinking modes and thinking budget configuration.

Overview

Extended thinking (also called "ultrathink") allows Claude to perform deeper reasoning before responding. This uses additional tokens for internal reasoning that improves quality on complex tasks.

Thinking Modes

Trigger Phrases (CLI)

Use these phrases in your prompts to activate thinking levels:

| Phrase | Thinking Tokens | Approx Cost | |--------|----------------|-------------| | think | ~4,000 tokens | ~$0.06 | | think hard / megathink | ~10,000 tokens | ~$0.15 | | ultrathink | ~32,000 tokens | ~$0.48 |

Example: "ultrathink about how to refactor the auth module"

Toggle Shortcut

• Option+T — Toggle extended thinking on/off in interactive mode

Environment Variable

CLAUDE_CODE_EFFORT_LEVEL=high   # low, medium, high

Default Thinking

Standard reasoning — Claude thinks internally as needed.

Extended Thinking

More deliberate reasoning with configurable thinking budget.

Ultrathink

Maximum reasoning depth for the most complex problems. Best for architecture decisions, debugging complex issues, and novel problem-solving.

Configuration

Via API (Claude Agent 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: 10000  // Tokens allocated for thinking
  },
  messages: [
    { role: "user", content: "Solve this complex architecture problem..." }
  ]
});

// Access thinking content
for (const block of response.content) {
  if (block.type === "thinking") {
    console.log("Thinking:", block.thinking);
  } else if (block.type === "text") {
    console.log("Response:", block.text);
  }
}

Budget Tokens

| Budget | Use Case | |--------|----------| | 2,000-5,000 | Simple analysis, code review | | 5,000-10,000 | Architecture decisions, debugging | | 10,000-20,000 | Complex multi-step reasoning | | 20,000+ | Deep research, novel problem solving |

Streaming with Thinking

const stream = await client.messages.stream({
  model: "claude-sonnet-4-6",
  max_tokens: 16000,
  thinking: {
    type: "enabled",
    budget_tokens: 10000
  },
  messages: [{ role: "user", content: "..." }]
});

for await (const event of stream) {
  if (event.type === "content_block_start") {
    if (event.content_block.type === "thinking") {
      console.log("--- Thinking started ---");
    }
  }
  if (event.type === "content_block_delta") {
    if (event.delta.type === "thinking_delta") {
      process.stdout.write(event.delta.thinking);
    } else if (event.delta.type === "text_delta") {
      process.stdout.write(event.delta.text);
    }
  }
}

In Claude Code

Within Claude Code sessions, extended thinking is managed automatically based on:

1. Model capabilities
2. Task complexity
3. Available token budget

When Extended Thinking Activates

• Complex architectural decisions
• Multi-step debugging
• Code review requiring deep analysis
• Planning with many constraints
• Novel problem-solving

Model Support

| Model | Extended Thinking | |-------|------------------| | Claude Opus 4.6 | Full support | | Claude Sonnet 4.6 | Full support | | Claude Haiku 4.5 | Limited support |

Best Practices

1. Use for complex tasks — Don't waste thinking budget on simple questions
2. Set appropriate budgets — More isn't always better
3. Monitor costs — Thinking tokens count toward usage
4. Combine with planning — Extended thinking + plan mode for architecture
5. Review thinking output — Verify the reasoning, not just the conclusion

Cost Implications

• Thinking tokens are billed as output tokens
• Higher thinking budgets increase cost per request
• Use /cost to monitor thinking token usage
• Consider model routing: use Haiku for simple, Opus for complex

Thinking in Multi-Turn Conversations

• Thinking context carries across turns
• Each turn can have its own thinking budget
• Claude references previous thinking implicitly
• /compact preserves thinking conclusions but drops raw thinking

API Response Structure

{
  "content": [
    {
      "type": "thinking",
      "thinking": "Let me analyze this step by step...\n1. First consideration...\n2. Second consideration..."
    },
    {
      "type": "text",
      "text": "Based on my analysis, here's what I recommend..."
    }
  ],
  "usage": {
    "input_tokens": 500,
    "output_tokens": 3000,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0
  }
}