Claude Code Advisor:主模型干杂活,关键决策再请 Opus 顾问
Claude Code 的 Advisor 允许用 Sonnet/Haiku 做常规读写与测试,在选型、反复报错与收工前把完整会话交给 Opus 顾问;官方测例显示 Haiku+Opus 顾问在 BrowseComp 上明显高于单用 Haiku,且目前仅 Anthropic API。
Anthropic 在 Claude Code 文档中提供实验性 Advisor(顾问) 能力:不必全程使用最贵模型,而是让较便宜的主模型(Sonnet 或 Haiku)负责读文件、跑测试等常规步骤,只在“真正决定走向”的节点再请更强的顾问模型(例如 Opus)介入。
启用方式包括会话内 /advisor opus,或以 claude --advisor opus 启动。顾问会阅读完整会话记录(含每次工具调用与结果),再在例如选定实现路径之前、同一错误反复出现时、或宣告任务完成前给出意见。帖文强调:中途开关顾问不会使当前提示缓存失效,便于在长重构里按需加码。
Anthropic 自测数据(帖文转述):Haiku 搭配 Opus 顾问在 BrowseComp 上取得 41.2%,对比单用 Haiku 的 19.7%;Sonnet 搭配 Opus 顾问在 SWE-bench Multilingual 上比单用 Sonnet 高约 2.7 分,同时每任务成本约低 11.9%。局限同样写明:功能仍标为实验性,且目前仅 Anthropic API,尚未覆盖 Bedrock 等其它云托管入口。
flowchart TD
A[主模型 Sonnet/Haiku<br/>读写文件/跑测试] --> B{决策点?}
B -->|否| A
B -->|选型/反复报错/收工前| C[Advisor Opus<br/>读完整 transcript]
C --> D[调整方案或继续]
D --> A
对软件研发团队,这意味着可以把“按决策路由模型”写进本队 Claude Code 工作流:长任务默认省成本,关键分叉再加顾问;上线前仍应用本仓库真实任务复测缓存命中与顾问触发频率,并确认账号是否走 Anthropic API。
Claude Code advisor: route frontier model only to decision turns
Your coding agent doesn't need the most expensive model on every turn. Just on the few turns that decide the outcome.
That's the idea behind Claude Code's advisor tool:
→ A cheaper main model (Sonnet or Haiku) does the routine work → A stronger advisor (Opus) gets consulted at decision points: before picking an approach, when an error keeps coming back, before calling a task done → The advisor reads the full transcript, every tool call and result → Turn it on with /advisor opus, or launch with claude --advisor opus → Toggling it mid-session doesn't invalidate your prompt cache
In Anthropic's own tests, Haiku with an Opus advisor scored 41.2% on BrowseComp vs 19.7% alone. Sonnet with an Opus advisor scored 2.7 points higher on SWE-bench Multilingual at 11.9% lower cost per task.
My take: most agent turns are reading files and running tests. Paying frontier prices for those is the expensive habit. Route by decision, not by task.
Still experimental, and Anthropic API only for now (not Bedrock or the other clouds).
Bookmark this for your next long refactor 👇 https://code.claude.com/docs/en/advisor