latestaiagents/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

Do not feed thinking blocks back to the user as-is in production — they're not polished prose. And do not modify them before passing back in multi-turn (signature validation will fail).

Interleaved Thinking with Tools

With the interleaved-thinking-2025-05-14 beta, the model thinks between tool calls — reasoning about each tool result before picking the next:

const response = await client.messages.create(
  {
    model: "claude-sonnet-4-6",
    max_tokens: 16_000,
    thinking: { type: "enabled", budget_tokens: 10_000 },
    tools: [searchTool, fetchTool, summarizeTool],
    messages: [{ role: "user", content: "Research X and write a brief." }],
  },
  { headers: { "anthropic-beta": "interleaved-thinking-2025-05-14" } },
);

Without interleaved thinking, the model only thinks once at the start. With it, the model can reassess after every tool result — critical for agents that operate under uncertainty.

Multi-Turn Conversations

When continuing a conversation that included thinking, pass the assistant's full message back unchanged (including thinking blocks):

messages.push({ role: "assistant", content: response.content });
messages.push({ role: "user", content: "Great. Now prove the converse." });

const next = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 16_000,
  thinking: { type: "enabled", budget_tokens: 10_000 },
  messages,
});

Thinking blocks carry signatures that the API validates. Reordering or editing them breaks the request.

Cost

Thinking tokens are billed as output tokens. A call with 10K thinking + 2K answer costs 12K output tokens.

Rough rule: thinking doubles-to-triples the cost of a reasoning-heavy call. Confirm it's worth it by A/B testing against no-thinking.

Streaming

const stream = client.messages.stream({
  model: "claude-sonnet-4-6",
  max_tokens: 16_000,
  thinking: { type: "enabled", budget_tokens: 10_000 },
  messages: [...],
});

for await (const event of stream) {
  if (event.type === "content_block_delta" && event.delta.type === "thinking_delta") {
    // show a "thinking..." spinner or subtle text
  } else if (event.type === "content_block_delta" && event.delta.type === "text_delta") {
    process.stdout.write(event.delta.text);
  }
}

In UX, show a distinct "thinking" indicator, then switch to streaming the answer.

Anti-Patterns

1. Thinking on every request — massive cost increase with no quality gain for easy tasks
2. Tiny budget on hard tasks — 1000 tokens isn't enough for real reasoning; model truncates
3. Modifying thinking blocks between turns — breaks signature; API rejects
4. Showing raw thinking to end users in production — messy, not polished

Best Practices

1. Gate thinking on task complexity — classify the query first, enable thinking only for hard ones
2. Start budget at 5K; measure quality and cost; adjust
3. Enable interleaved thinking for any agent with > 3 tools
4. Pass thinking blocks back unmodified in multi-turn
5. Stream the response so users see progress during long thinks
6. Log thinking tokens used per request — detect when the model is consistently maxing out budget