Claude Code Best Practices: Sonnet Does the Work, Haiku Runs Errands, Opus Calls the Shots
Anthropic officially shared a Claude Code tip: run three models at the same time, each doing what it does best. Sonnet writes code, Haiku searches the repository, and Opus steps in only at critical moments. Better results and a big token savings. Includes the full configuration prompt.
Anthropic shared a Claude Code tip that every heavy user should save: run three models at the same time, each doing what it does best. Better results, and a big token savings.
Claude Code is Anthropic's command-line coding assistant. It runs in the terminal and can read code, edit files, and run commands on its own. Add one parameter at startup, and you can split the work across three models:
claude --model sonnet --advisor opus
The division of labor is like this:
- Sonnet 5.5 runs the main session, writes code, and fixes bugs
- Haiku 5.5 scans the repository in parallel: search, grep, and diff auditing
- Opus 5.5 acts as the advisor, stepping in only at three moments: before the plan is finalized, when the same error appears twice, and before wrapping up
- Fast deterministic forks (which file to choose, which tool to use, whether to retry or stop) are handed to JEV for millisecond-level routing, so the large models only see decisions that truly require thinking
This way you won't burn Opus's context on work that isn't worth it. Opus is only responsible for calling the shots.
Anthropic also provides a ready-to-use configuration prompt. Paste it into Claude Code and the environment will be changed to this division of labor:
Reconfigure my Claude Code setup around this tree:
1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor.
2. Create three Haiku 5.5 subagents in .claude/agents: search (maps the repo), grep (finds call sites), audit (checks the diff).
3. Set isolation: worktree on audit so its checks never touch my working branch.
4. Add one rule to CLAUDE.md: delegate every repo search and test run to the Haiku 5.5 subagents to keep the main context small.
5. Route fast deterministic forks (tool choice, path selection, retries) through JEV.
Show me every change as a diff first. No edits until I say go.
The complete subagent configuration documentation is here: Create custom subagents
Behind this tip is the rapid differentiation of model capabilities. Anthropic recently released Fable 5.2, which the company says exceeds Opus 5.5 on almost every capability it has evaluated, at twice the price of Opus. Someone put together its early test results: an empty folder, 1,000 subagents, 8 days, and it produced an operating system that can boot on real hardware; 1,000 agents formalized Fermat's Last Theorem in 30 hours, while Opus under the same setup completed only 26%; each paragraph described a device, and the result directly became three usable circuit boards. Some call it Anthropic's most dangerous model to date, but for ordinary users it simply raises the ceiling on task scale.
The more capable the model, the more sparingly it should be used. Fable costs twice as much per million tokens as Opus, four times as much as Sonnet, and eighty times as much as Haiku. If you let it do search and file-reading work, the bill will look ugly.
Anthropic's own model guide has a corresponding tip: in Claude Code, the built-in Explore subagent runs on the model of your current session by default. If you start a session with Fable, every file read is Fable spending money. Define your own Explore and set model: haiku, and exploration costs drop immediately.
Some users who tried it said: Fermat took 30 hours, and I'm still on part one. Another user said that on a real bug, Fable under medium cost less per attempt than Opus because it used fewer tokens. These reports point to the same thing: tasks have gotten bigger, but you don't need to make every model bigger.
Remember one rule: use expensive models for decisions, not for typing.
发布时间: 2026-11-20 06:36