Skip to main content
This page describes Anthropic’s Extended Thinking feature. To learn more about Anthropic models, see Anthropic Overview.

​
Quick Start

Enable deep reasoning for complex problems by allocating token budget for internal analysis before generating responses. 
curl -X POST https://api.orq.ai/v2/proxy/chat/completions \
  -H "Authorization: Bearer $ORQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-opus-4-5-20251101",
    "messages": [
      {
        "role": "user",
        "content": "Design a distributed rate limiting system for 1M requests/second"
      }
    ],
    "thinking": {
      "type": "enabled",
      "budget_tokens": 8000
    },
    "max_tokens": 2048
  }'

​
Multi-turn Conversations

When building chat applications with extended thinking, properly map the reasoning content with its signature for conversation history: 
const messages = [
  { role: "user", content: "Design a rate limiting system" }
];

const response = await openai.chat.completions.create({
  model: "anthropic/claude-opus-4-5-20251101",
  messages,
  thinking: { type: "enabled", budget_tokens: 8000 },
  max_tokens: 2048
});

// Map response to assistant message
const msg = response.choices[0].message;
const contentParts = [];

// Add reasoning if present (flat field in response)
if (msg.reasoning) {
  contentParts.push({
    type: "reasoning",
    reasoning: msg.reasoning,
    signature: msg.reasoning_signature
  });
}

// Add redacted reasoning if present (flat field in response)
if (msg.redacted_reasoning) {
  contentParts.push({
    type: "redacted_reasoning",
    data: msg.redacted_reasoning
  });
}

// Add main text content
if (msg.content) {
  contentParts.push({
    type: "text",
    text: msg.content
  });
}

const assistantMessage = {
  role: "assistant",
  content: contentParts
};

// Continue conversation
messages.push(assistantMessage);
messages.push({ role: "user", content: "How would you handle 10M req/s?" });

const followUp = await openai.chat.completions.create({
  model: "anthropic/claude-opus-4-5-20251101",
  messages,
  thinking: { type: "enabled", budget_tokens: 8000 },
  max_tokens: 2048
});
Important: Always include the signature field when passing reasoning content back to the API. The signature cryptographically verifies the reasoning was generated by the model and is required for multi-turn conversations.
If the model’s reasoning is flagged by safety systems, it will be returned as redacted_reasoning (encrypted). Pass this back with type: "redacted_reasoning" so the model can decrypt and continue without losing context.

​
Configuration

ParameterTypeRequiredDescription
thinking.type"enabled"YesEnables extended thinking
thinking.budget_tokensnumberYesTokens allocated for thinking (1024-64000)
Budget recommendations:
  • 2000-5000: Code reviews, moderate analysis
  • 5000-10000: System design, strategic planning
  • 10000+: Research, complex multi-step reasoning
Important: Budget tokens are billed as output tokens and count toward total usage.

​
Supported Models

Extended thinking is available on Claude Opus 4.5 (recommended), Sonnet 4.5, and newer models. For the complete list, see Anthropic’s documentation. Provider availability: anthropic/, aws/, google/

​
When to Use

Best for:
  • System architecture and design decisions
  • Code optimization and performance analysis
  • Strategic planning with constraints
  • Complex mathematical or logical problems
  • Multi-step analysis requiring deep reasoning
Skip for:
  • Simple queries or factual questions
  • Time-sensitive operations (adds 2-4s per 1000 tokens)
  • High-volume, low-complexity tasks
  • Budget-constrained scenarios

​
Best Practices

Choose budget based on complexity: 
const getBudget = (taskType: string) => ({
  code_review: 4000,
  system_design: 8000,
  research: 12000,
}[taskType] || 5000);
Use clear, structured prompts: 
// Good: Encourages systematic thinking
"Analyze this architecture design, considering scalability, cost, and maintainability"

// Less effective: Vague request
"Look at this design"
Combine with prompt caching for repeated contexts: 
const response = await openai.chat.completions.create({
  model: "anthropic/claude-opus-4-5-20251101",
  messages: [
    {
      role: "system",
      content: [{
        type: "text",
        text: "You are a system architect...", // Cache this
        cache_control: { type: "ephemeral" }
      }]
    },
    { role: "user", content: "Design a notification system" }
  ],
  thinking: { type: "enabled", budget_tokens: 8000 }
});