Extended Thinking

Extended Thinking (Ultrathink) Skill

Enable Claude's extended thinking capabilities for complex reasoning tasks that benefit from internal deliberation before responding.

When to Use

• Complex problem solving requiring multi-step reasoning
• Code architecture decisions with multiple trade-offs
• Debugging complex issues needing systematic analysis
• Strategic planning with many variables
• Mathematical or logical proofs
• Security analysis requiring threat modeling
• Performance optimization with multiple factors

Supported Models

API Configuration

Adaptive Thinking — current models (Python)

On current models (Opus 4.6+, Sonnet 5, Fable 5), adaptive thinking replaces manual budgets: Claude decides when and how much to think, and output_config.effort controls depth.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    output_config={"effort": "high"},  # low | medium | high | xhigh | max
    messages=[{
        "role": "user",
        "content": "Analyze this complex architecture decision..."
    }]
)

# Access thinking and response
for block in response.content:
    if block.type == "thinking":
        print(f"Thinking: {block.thinking}")
    elif block.type == "text":
        print(f"Response: {block.text}")

TypeScript Configuration

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

const client = new Anthropic();

const response = await client.messages.create({
  model: "claude-sonnet-5",
  max_tokens: 16000,
  thinking: { type: "adaptive" },
  output_config: { effort: "high" },
  messages: [
    {
      role: "user",
      content: "Analyze this complex architecture decision...",
    },
  ],
});

Manual Budgets — older models only (budget_tokens)

On older models (Sonnet 4.5 / Haiku 4.5 era), extended thinking is enabled with a fixed token budget. budget_tokens must be >= 1,024 and < max_tokens. Do not use this shape on current models — budget_tokens is removed on Opus 4.7+, Sonnet 5, and Fable 5 and returns a 400 error (deprecated but still functional on Opus 4.6 / Sonnet 4.6).

response = client.messages.create(
    model="claude-sonnet-4-5",  # older model — manual budget still applies
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000  # Minimum 1,024
    },
    messages=[{"role": "user", "content": "..."}]
)

Streaming (Required for large max_tokens)

with client.messages.stream(
    model="claude-sonnet-5",
    max_tokens=32000,
    thinking={"type": "adaptive"},
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": prompt}]
) as stream:
    for event in stream:
        if event.type == "content_block_delta":
            if hasattr(event.delta, "thinking"):
                print(event.delta.thinking, end="", flush=True)
            elif hasattr(event.delta, "text"):
                print(event.delta.text, end="", flush=True)

Thinking Depth Recommendations

Current models — effort levels (output_config: {"effort": ...})

Older models — budget_tokens

Tool Use with Extended Thinking

CRITICAL: When using tools with extended thinking, you MUST preserve thinking blocks in the conversation history.

# Initial request with thinking
response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    tools=[{
        "name": "analyze_code",
        "description": "Analyze code for issues",
        "input_schema": {
            "type": "object",
            "properties": {"code": {"type": "string"}},
            "required": ["code"]
        }
    }],
    messages=[{"role": "user", "content": "Analyze this code..."}]
)

# MUST include ALL content blocks including thinking
tool_use_block = next(b for b in response.content if b.type == "tool_use")
tool_result = execute_tool(tool_use_block)

# Continue with thinking blocks preserved
follow_up = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    tools=[...],
    messages=[
        {"role": "user", "content": "Analyze this code..."},
        {"role": "assistant", "content": response.content},  # Includes thinking!
        {"role": "user", "content": [{
            "type": "tool_result",
            "tool_use_id": tool_use_block.id,
            "content": tool_result
        }]}
    ]
)

Interleaved Thinking

On current models (Opus 4.6+, Sonnet 5, Fable 5), adaptive thinking automatically interleaves thinking between tool calls — no beta header needed:

response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    messages=[...]
)

On older Claude 4 models using manual budget_tokens, interleaved thinking requires the interleaved-thinking-2025-05-14 beta header.

Constraints

• Current models (Opus 4.7+, Sonnet 5, Fable 5): budget_tokens, temperature, top_p, and top_k are all removed — sending any of them returns 400
• Older models: minimum budget_tokens is 1,024 and must be < max_tokens; temperature must be 1 (default); top_k cannot be used with extended thinking
• Maximum output: 128k tokens (thinking + response); Haiku 4.5 caps at 64k
• Streaming required: For large max_tokens (above ~16k, to avoid SDK HTTP timeouts)
• System prompts: Fully compatible

Best Practices

1. Start conservative: Begin with lower budgets, increase as needed
2. Monitor actual usage: Track thinking_tokens in response usage
3. Use streaming: For better UX and larger outputs
4. Preserve thinking blocks: Critical for multi-turn tool use
5. Batch for heavy workloads: Use batch API for budgets > 32k tokens
6. Match budget to task: Don't over-allocate for simple tasks

Integration with Claude Code

When using Claude Code CLI with extended thinking models:

# The CLI automatically handles extended thinking for supported models
# Use opus or sonnet models for complex tasks

claude --model claude-opus-4-8 "Analyze this codebase architecture"

See Also

• [[complex-reasoning]] - Multi-step reasoning patterns
• [[deep-analysis]] - Analytical thinking templates
• [[llm-integration]] - General LLM API patterns