Claude Code 三模型分工：Sonnet 写码、Haiku 扫描、Opus 审核

Bober_smart @Bober_smart原推发布于 2026.10.26

在 X 上打开原推

THIS IS AN ABSOLUTE GOLDMINE  The team at Anthropic just revealed one of the most effective techniques for Claude Code The core concept: run Sonnet, Haiku 5.5, and Opus 5.5 simultaneously, letting each model focus strictly on what it does best You get higher quality output while dramatically reducing token consumption, since every specific job gets routed to the optimal model To spin up Claude Code with this setup, simply execute:  ⁠claude --model sonnet --advisor opus⁠  Here is how the workload is split:  > Sonnet 5.5 handles the primary session, writes code, and applies bug fixes > Haiku 5.5 scans the codebase in parallel: performing searches, running grep queries, exploring docs, and auditing diffs > Opus 5.5 acts as an on-demand advisor for critical decisions and approves final code diffs > You stop burning expensive Opus context on routine tasks that don't require high-level reasoning Opus intervenes exclusively during three specific scenarios:  > Right before locking in an execution plan > Whenever the exact same error triggers twice in a row > Prior to marking a task as complete Sonnet builds. Haiku scans and explores. Opus reviews JEV engineering applies the exact same principle one layer deeper: routine execution forks that don't need high-level reasoning (determining which file to inspect, which tool to trigger, or whether to retry or stop) are processed by JEV in sub-milliseconds, leaving big models to handle only complex decision splits The full architectural tree: Sonnet 5.5 manages the primary session and modifies code Search maps out the repository structure via Haiku 5.5 Grep locates call references using Haiku 5.5 Docs fetches API specifications through Haiku 5.5 Opus 5.5 serves as an advisor, reading the complete conversation history whenever consulted JEV routes lightning-fast micro-forks in milliseconds Feed this tree structure alongside the prompt below into Claude Code Reconfigure my Claude Code workspace based on this architecture: 1. Run the primary session utilizing Sonnet 5.5, with Opus 5.5 designated as the advisor 2. Set up three read-only Haiku 5.5 subagents: search (maps out the repo), grep (locates call sites), and docs (retrieves API documentation) 3. Insert a single rule into CLAUDE.md: consult the advisor prior to committing to a plan, after encountering an identical error twice, and before marking a task as done 4. Direct rapid, deterministic decisions (tool selections, file paths, retries) through JEV. Show all modifications as a diff first. Do not execute any edits until I grant permission Bookmark this setup: it will become a daily driver in your workflow

这条是引用着另一条写的：x.com/bober_smart/status/2078784709253841039