Thinking Mode

Claude Thinking mode lets the model perform internal reasoning before answering. It significantly improves accuracy on complex reasoning tasks such as math, code debugging, and multi-step planning, at the cost of more output tokens and latency.

Two Ways to Enable It

Method 1: Model Name Suffix (Recommended, available with any protocol)

Use the -thinking tier model directly, with no other parameters required:

• claude-opus-4-6-thinking
• claude-sonnet-4-6-thinking

Method 2: Native thinking parameter (Claude native API)

{
  "model": "claude-sonnet-5",
  "max_tokens": 16000,
  "thinking": { "type": "enabled", "budget_tokens": 8000 },
  "messages": [{ "role": "user", "content": "Prove: among any 6 people, there must be 3 who all know each other or 3 who all do not know each other" }]
}

budget_tokens is the token budget for the thinking process, and must be less than max_tokens.

Output Parsing

In the response content, thinking and answer are returned in separate blocks:

{
  "content": [
    { "type": "thinking", "thinking": "First, convert the problem into graph coloring..." },
    { "type": "text", "text": "This is the classic Ramsey number problem..." }
  ]
}

• When displayed in the UI, the thinking block is collapsed or hidden, and only the text block is treated as the answer;
• Thinking tokens are billed as output tokens, so setting the budget too high will significantly increase cost;
• In streaming mode, thinking is sent incrementally via thinking_delta; see Streaming and Non-streaming Responses.

Selection Recommendations

• Use the standard tier for ordinary tasks; use the thinking tier for grading, auditing, and tackling hard problems;
• For comparisons with GPT/Gemini reasoning tiers, see Reasoning Model Output.