Three-layer Claude Code setup cuts daily cost from $224 to $20
Splitting a Claude Code session into Sonnet as lead, Haiku subagents for reads and Opus as advisor, with Jev handling mechanical decisions in 16ms, cut a day's bill by 91%.
ANTHROPIC SHIPPED HAIKU 5.5 YESTERDAY. MY CLAUDE CODE DAY WENT FROM $224.90 TO $20.16. Claude Code subagents bill at Opus rates by default. Every helper you spawn reads your repo at $4 in and $20 out per million tokens. Haiku 5.5 does the same reads for $0.10 in and $0.50 out. That's a 40x gap on every subagent. So I split the session into three layers. → Sonnet 5.5 drives the main session on high effort. It writes the diffs and runs the tests. → Haiku 5.5 subagents swarm the repo in parallel. File search, docs lookup, log summaries. Read only. → Opus 5.5 stays on call as the advisor and wakes up at three moments. → Before a plan locks. Does it miss an auth rule or a schema contract? → When the same test breaks twice. Are we fixing the cause or digging a deeper hole? → Before anything gets called done. Did the diff break something nobody asked about? Opus reviews. Sonnet builds. Haiku swarms. Haiku never touches a file. It scores 39.2% on Terminal-Bench 4.0, Sonnet scores 70.6%, so Haiku gets the lookups and Sonnet keeps the pen. One read is about 30K tokens in and 2K out. That's $0.004 on Haiku and $0.16 on Opus. Yesterday the swarm did 1,240 reads for $4.96. On Opus the same reads would have cost $198.40. $4.96 Haiku + $11.30 Sonnet + $3.90 Opus = $20.16 for the whole day. The same day on Opus alone is $224.90. 91% of the bill is gone. One layer down, Jev makes the mechanical calls. Which file to open, which tool to run, retry or stop. Those answers come back in 16ms and never wake a frontier model. My Max 20x plan now includes $200 of API credit a month, and it works on Haiku. That covers 9 days of this setup. The first morning went wrong. I opened the subagent transcripts and 214 of 230 had run on Opus. My Claude Code was a few builds behind, and model: haiku got ignored. Those 214 reads cost $34.24. On Haiku they would have been 86 cents. Check that the routing took. Open any transcript in the subagents folder under ~/.claude/projects/. The first assistant message tells you which model ran. Keep every read under 100K tokens. Past that line, Haiku costs 5x more. The full setup: > Sonnet 5.5 lead on high effort, edits and tests. > Haiku 5.5 subagents on low effort, no Edit or Write tools. > Opus 5.5 advisor on call, three checkpoints written into CLAUDE.md. > Jev on routine routing calls. > A transcript check after every first run. Paste this into Claude Code ↓ "Rebuild my Claude Code setup into three layers. 1. Check that Claude Code is up to date, so the haiku alias resolves to Haiku 5.5. If it isn't, stop and tell me to run claude update. 2. Create or update subagents in .claude/agents/. Explore and researcher get model: haiku, effort: low, and only the Read, Grep and Glob tools. Worker gets model: sonnet, effort: medium. 3. Set the main session to sonnet on high effort, with opus as the advisor. 4. Add one line to CLAUDE.md: consult the advisor before a big plan, when the same error shows up twice, and before marking a task done. 5. After the first run, open the subagent transcripts and report which model each one actually used. Show every change as a diff first. Wait for my go-ahead before applying anything." 1,240 reads. $4.96. Zero edits by Haiku.