Claude 4.6 モデル選択 実践ガイド — Sonnet vs Opus、コスト・性能・Claude Codeでの使い分けを徹底解説

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

const client = new Anthropic();

// Sonnet 4.6（コスト重視の日常タスク）
const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 4096,
  messages: [{ role: "user", content: "このコードをレビューしてください" }],
});

// Opus 4.6（複雑な推論が必要な場面）
const response = await client.messages.create({
  model: "claude-opus-4-6",
  max_tokens: 4096,
  messages: [{ role: "user", content: "このアーキテクチャの改善案を提案してください" }],
});

Error: This model does not support assistant message prefill. 
The conversation must end with a user message.

// ❌ 4.6ではこれが400エラーになる
const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 4096,
  messages: [
    { role: "user", content: "売上データを分析してください" },
    { role: "assistant", content: '{"analysis":' }, // Prefill
  ],
});

// ✅ system promptで出力形式を指示
const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 4096,
  system: "必ずJSON形式で回答してください。キー: analysis, summary, recommendations",
  messages: [
    { role: "user", content: "売上データを分析してください" },
  ],
});

// ❌ 非推奨（まだ動くが将来削除予定）
const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 16000,
  thinking: { type: "enabled", budget_tokens: 10000 },
  messages: [{ role: "user", content: "..." }],
});

// ✅ Adaptive Thinking（推奨）
const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 16000,
  thinking: { type: "adaptive" },
  // effort: "medium" | "high" | "max"（省略時はhigh）
  messages: [{ role: "user", content: "..." }],
});

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

type ModelTier = "haiku" | "sonnet" | "opus";

interface RouterConfig {
  tokenThreshold: { sonnet: number; opus: number };
  complexityKeywords: string[];
}

const DEFAULT_CONFIG: RouterConfig = {
  tokenThreshold: { sonnet: 500, opus: 5000 },
  complexityKeywords: [
    "アーキテクチャ", "設計", "リファクタ", "セキュリティ監査",
    "パフォーマンス最適化", "architecture", "refactor", "security audit",
  ],
};

const MODEL_MAP: Record<ModelTier, { id: string; inputCost: number; outputCost: number }> = {
  haiku: { id: "claude-haiku-4-5-20251001", inputCost: 1, outputCost: 5 },
  sonnet: { id: "claude-sonnet-4-6", inputCost: 3, outputCost: 15 },
  opus: { id: "claude-opus-4-6", inputCost: 5, outputCost: 25 },
};

function selectModel(prompt: string, config = DEFAULT_CONFIG): ModelTier {
  const estimatedTokens = Math.ceil(prompt.length / 3); // 粗い推定
  const hasComplexity = config.complexityKeywords.some((kw) =>
    prompt.toLowerCase().includes(kw.toLowerCase())
  );

  if (hasComplexity || estimatedTokens > config.tokenThreshold.opus) return "opus";
  if (estimatedTokens > config.tokenThreshold.sonnet) return "sonnet";
  return "haiku";
}

// 使用例
async function routedRequest(prompt: string) {
  const client = new Anthropic();
  const tier = selectModel(prompt);
  const model = MODEL_MAP[tier];

  console.log(`[Router] Selected: ${tier} (${model.id})`);

  const response = await client.messages.create({
    model: model.id,
    max_tokens: 4096,
    messages: [{ role: "user", content: prompt }],
  });

  const inputTokens = response.usage.input_tokens;
  const outputTokens = response.usage.output_tokens;
  const cost = (inputTokens * model.inputCost + outputTokens * model.outputCost) / 1_000_000;
  console.log(`[Router] Cost: $${cost.toFixed(6)} (${inputTokens} in / ${outputTokens} out)`);

  return response;
}

// Prompt Cachingの基本パターン
const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 4096,
  system: [
    {
      type: "text",
      text: "あなたはコードレビューの専門家です。以下のコーディング規約に従ってレビューしてください...(長い規約テキスト)",
      cache_control: { type: "ephemeral" }, // 5分間キャッシュ
    },
  ],
  messages: [{ role: "user", content: "このPRをレビューしてください: ..." }],
});

タスクの判断フロー:

  入力が短い＆単純な質問/補完？
    → Yes: Haiku 4.5（最安・最速）
    → No: ↓

  複雑な推論・大規模コード解析・高精度が必須？
    → Yes: Opus 4.6（effort: "max"）
    → No: ↓

  それ以外の開発タスク全般
    → Sonnet 4.6（effort: "medium" or "high"）