Claude Code 分层模型配置：Sonnet 主跑、Haiku 检索、Opus 复核

Shelpid.WI3M @Shelpid_WI3M原推发布于 2026.10.27

在 X 上打开原推

STOP PAYING OPUS PRICES FOR WORK HAIKU CAN DO. Your strongest model doesn’t need to handle every search, file lookup, and routine edit. Give each model a specific job: → Sonnet 5.5 runs the main session, writes code, and fixes bugs. → Haiku 5.5 handles focused, read-only searches across code and documentation. → Opus 5.5 advises on difficult decisions and reviews the final changes. Sonnet builds. Haiku investigates. Opus reviews. The goal: spend less expensive model context on routine work while keeping deeper reasoning available where it matters. Consult Opus at three checkpoints: • Before committing to an implementation plan. • After the same error occurs twice. • Before declaring the task complete. For workflows using JEV, put simple, deterministic routing decisions there: tool selection, file selection, and bounded retries. Paste this into Claude Code: “Inspect my installed Claude Code version and available integrations, then propose this configuration: 1. Use Sonnet 5.5 for the main session and Opus 5.5 as an advisor. 2. Create three read-only Haiku 5.5 subagents: search for repository structure, grep for call sites, and docs for API documentation. 3. Add a CLAUDE.md rule to consult the advisor before finalizing a plan, after two identical errors, and before completion. 4. If JEV is installed and supports it, use it for deterministic routing decisions. 5. Verify the model names and configuration options. Flag anything unsupported instead of inventing settings. 6. Show the proposed changes as a diff. Wait for my approval before editing.” Start with a clear division of work. Measure cost and output quality before expanding it. Bookmark this for your next Claude Code session.