Claude Adaptive Thinking実践ガイド — budget_tokensからの移行と4段階effort制御で実現するコスト最適な推論戦略

from anthropic import Anthropic
client = Anthropic()

# 1. Adaptive Thinking（推奨）
response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=16000,
    thinking={
        "type": "adaptive",    # Claudeが思考量を自律判断
    },
    output_config={
        "effort": "high",      # 思考の深さをガイド
    },
    messages=[{"role": "user", "content": "この関数のバグを特定してください"}],
)

# 2. Manual Extended Thinking（非推奨・将来削除予定）
response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000,  # 固定トークン予算
    },
    messages=[{"role": "user", "content": "この関数のバグを特定してください"}],
)

# 3. Thinking無効（思考なし）
response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=16000,
    messages=[{"role": "user", "content": "こんにちは"}],
)

response = client.beta.messages.create(
    model="claude-opus-4-5-20250415",
    max_tokens=16000,
    betas=["interleaved-thinking-2025-05-14"],
    thinking={
        "type": "enabled",
        "budget_tokens": 10000,
    },
    messages=[{"role": "user", "content": prompt}],
)

response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=16000,
    thinking={
        "type": "adaptive",
    },
    output_config={
        "effort": "high",
    },
    messages=[{"role": "user", "content": prompt}],
)

import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();

// ユースケースに応じてeffortを動的に選択
function getEffortForTask(taskType: string): "low" | "medium" | "high" | "max" {
  switch (taskType) {
    case "classification":
    case "lookup":
      return "low";
    case "chat":
    case "code_generation":
      return "medium";
    case "complex_reasoning":
    case "agent":
      return "high";
    case "critical_analysis":
      return "max";  // Opus 4.6専用
    default:
      return "high";
  }
}

const effort = getEffortForTask("code_generation");

const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 8000,
  thinking: { type: "adaptive" },
  output_config: { effort },
  messages: [{ role: "user", content: "FizzBuzzを実装してください" }],
});

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

const client = new Anthropic();

type Phase = "gather" | "analyze" | "decide";

function getEffortForPhase(phase: Phase, model: string) {
  const effortMap: Record<Phase, "low" | "medium" | "high" | "max"> = {
    gather: "low",       // 情報収集は高速に
    analyze: "high",     // 分析はじっくり
    decide: model === "claude-opus-4-6" ? "max" : "high", // 最終判断は最高精度
  };
  return effortMap[phase];
}

async function agentStep(
  phase: Phase,
  messages: Anthropic.Messages.MessageParam[],
  tools: Anthropic.Messages.Tool[],
) {
  const model = "claude-opus-4-6";
  const effort = getEffortForPhase(phase, model);

  const response = await client.messages.create({
    model,
    max_tokens: 16000,
    thinking: { type: "adaptive" },
    output_config: { effort },
    tools,
    messages,
  });

  // stop_reasonに応じたハンドリング
  if (response.stop_reason === "max_tokens") {
    console.warn("max_tokens到達: max_tokensを増やすかeffortを下げてください");
  }

  return response;
}

const stream = await client.messages.stream({
  model: "claude-opus-4-6",
  max_tokens: 16000,
  thinking: { type: "adaptive" },
  output_config: { effort: "high" },
  messages: [{ role: "user", content: "複雑なアルゴリズムを設計してください" }],
});

for await (const event of stream) {
  if (event.type === "content_block_delta") {
    if (event.delta.type === "thinking_delta") {
      process.stdout.write(`[thinking] ${event.delta.thinking}`);
    } else if (event.delta.type === "text_delta") {
      process.stdout.write(event.delta.text);
    }
  }
}

// stop_reason別のエラーハンドリング
function handleResponse(response: Anthropic.Messages.Message) {
  switch (response.stop_reason) {
    case "end_turn":
      // 正常完了
      break;
    case "max_tokens":
      // 出力がmax_tokensに到達。max_tokensを増やすかeffortを下げる
      console.warn("出力が途中で切れています。max_tokensの増加を検討してください");
      break;
    case "refusal":
      // 安全性フィルターによる拒否
      console.error("リクエストが拒否されました");
      break;
    case "tool_use":
      // ツールコール。ツールを実行して結果を返す
      break;
  }

  // model_context_window_exceededはAPIエラーとして返る（stop_reasonではない）
}