The Most Expensive Model Only Appears at Three Moments: Claude Code's Three-Model Division of Labor
Sonnet does the work, Haiku runs errands, Opus makes the final call. Anthropic's shared Claude Code three-model division of labor uses fewer tokens and gets better results.
Anthropic recently shared a Claude Code tip: run Sonnet 5.5, Haiku 5.5, and Opus 5.5 at the same time, with each model doing only what it does best. X user @xsmryze broke it down into a configuration you can copy directly, with better results and fewer tokens.
Startup command:
claude --model sonnet --advisor opus
Division of labor:
- Sonnet 5.5 runs the main session, writes code, and applies fixes
- Haiku 5.5 scans the repo in parallel: search, grep, diff audit—read-only operations with independent context
- Opus 5.5 serves as an online advisor, handles big decisions, and signs off on the final diff
Opus steps in only at three moments:
- Before the plan is locked
- When the same error appears twice
- Before wrapping up
Sonnet builds, Haiku searches, Opus reviews. Opus's context is no longer consumed by search logs and grep results.
One layer down is JEV routing. For forks that don't need thinking—which file to choose, which tool to use, whether to retry or stop—JEV handles them at millisecond speed, and the large model only sees the forks that truly require judgment.
Full structure:
- Sonnet 5.5: main session, coordinates all work
- search (Haiku 5.5): read-only, maps the repo structure
- grep (Haiku 5.5): finds call sites in an independent context
- audit (Haiku 5.5): checks the diff in a temporary worktree, without touching the working branch
- Opus 5.5: online advisor, reviews the aggregated diff before signing off
- JEV: millisecond-level routing for micro-forks
Throw this tree and this prompt together into Claude Code:
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.
After reading it, some developers said this is the right way to use Opus—don't waste the expensive model on simple tasks. Some admitted that all 12 agents in their app run the large model and they never thought about splitting them; maybe they should hand boring tickets to Haiku and see whether merge quality after review gets worse.
This hits on most people's default setup: if you have the strongest model, you let it do everything. Search, grep, reading logs—Opus isn't better at these jobs than Haiku, and the price difference is dozens of times. Separating repetitive work from judgment saves tokens and improves quality.
A developer added that if you run this setup with an API key, Claude Code maps models through environment variables: ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL; the Haiku slot also runs background tasks.
To configure sub-agents manually, the official docs are here: http://code.claude.com/docs/en/sub-agents
发布时间: 2026-11-22 20:50