extended-thinking

Extended Thinking

Extended thinking gives the model a scratchpad before the final answer. Pay for reasoning tokens, get deeper answers. Use it surgically, not everywhere.

When to Use

• Complex reasoning: math, proofs, multi-step logic
• Code generation where the model needs to plan before writing
• Agent planning: deciding which of many tools to call and in what order
• Debugging subtle issues: model can "think through" root causes
• Writing where structure and coherence matter more than speed

When NOT to Use

• Simple classification, extraction, or formatting — pure overhead
• Latency-sensitive paths — thinking adds 2-30 seconds
• Small prompts where the model gets it right zero-shot anyway
• High-volume batch tasks on a tight cost budget

Enabling Thinking

const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 16_000,
  thinking: {
    type: "enabled",
    budget_tokens: 10_000,
  },
  messages: [{ role: "user", content: "Prove that every prime > 3 is of the form 6k±1." }],
});

budget_tokens is the max thinking tokens. Model may use fewer. max_tokens must be > budget_tokens (thinking counts toward the total).

Budget Sizing

Task	Typical budget
Short multi-step reasoning	2,000-5,000
Code generation with planning	5,000-10,000
Complex math/proofs	10,000-32,000
Deep agent planning	10,000-20,000
Research synthesis	16,000-32,000

Start at 5,000 and measure. Bigger budget ≠ better answers past a point.

Inspecting the Thinking Trace

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);
  }
}

The thinking block reveals the model's reasoning. Useful for:

• Debugging why the model chose a wrong answer
• Surfacing rationale to power users (e.g., "show reasoning" toggle)
• Auditing agent decisions