Claude Code 用 guard 把粗活转给 Haiku 省钱

Sprytix @Sprytixl原推发布于 2026.10.23

在 X 上打开原推

THIS IS A PURE F*CKING GOLDMINE Anthropic just shared one of the best tricks for Claude Code. The idea: put a guard between the expensive model and the grunt work. Opus 5.5 asks to read 48 files, the guard says no and hands it to the swarm. Same session, two ways: all-Opus would have burned $3.41. With the swarm it came out at $0.59. That is $2.82 kept, 83% never burned. You just launch Claude Code like this: claude --model opus # agents: model: haiku here's the split: → Opus 5.5 plans and decides, 97k tokens for $0.52 → every read, grep and test request gets rerouted to Haiku 5.5 → 12 Haiku agents map files, find call sites and run tests in parallel → Opus comes back to 3 summaries at 6k tokens, never raw files the part everyone gets wrong: they let the expensive model do the reading. Opus asked for 48 files in one go. the guard caught it, sent it to the swarm, and that single reroute kept $0.27 on its own. Haiku is 40x cheaper per token. mapping 32 files, finding 25 call sites, running 212 tests, all of it came in under a cent per lane. one catch worth knowing: past 100k of context Haiku jumps to 5x its own price, so every agent is kept small on purpose. grunt work never runs on Opus. that is the whole rule. Opus thinks. Haiku reads. The guard decides which is which. JEV engineering is the same move one layer down: 1,420 forks this run - grunt or think, which agent, retry or stop - resolved in code at 0.80 sharpness, and only the split ones ever reach Opus the full tree > Opus 5.5 runs the main session, plans and reviews only > a guard sits between the brain and the tools and reroutes grunt work > 12 Haiku 5.5 agents fan out across read, grep and test, each kept under 100k > every agent returns a summary, never raw files or logs > Opus 5.5 reads 3 summaries at 6k tokens and signs off > JEV resolves 1,420 micro-forks in code and escalates only the split ones paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the main session on Opus 5.5 for planning, decisions and review only. 2. Add a guard: any read, grep or test request from the main session gets rerouted to a Haiku 5.5 agent instead. 3. Fan the grunt work across parallel lanes, each agent kept under 100k of context. 4. Agents return summaries only, raw files and logs never enter the main session. 5. Add one rule to CLAUDE.md: grunt work never runs on Opus, no exceptions. Route deterministic forks (grunt or think, agent choice, retries) through JEV. Show me every change as a diff first. No edits until I say go." ↳ http://code.claude.com/docs/en/model-config save it, it'll come in handy.

这条是引用着另一条写的：x.com/sprytixl/status/2107763347097534680