Opus plans, Haiku executes: Jev resolves 1,301 micro-forks in code

Sprytix @SprytixlOriginally posted 2026.10.22

Open original on X

THIS IS A PURE F*CKING GOLDMINE Anthropic just shared one of the best tricks for Claude Code. The idea: keep one expensive brain and make everything else cheap. Opus 5.5 thinks, Haiku 5.5 does the grunt work, and your bill drops 79%. Same task, two ways: all-Opus would have cost $1.13. This run cost $0.23. You just launch Claude Code like this: claude --model opus # agents: model: haiku here's the split: → Opus 5.5 plans, decides and reviews, and burns 40k tokens for $0.21 → 12 Haiku 5.5 agents fan out across read, grep and test lanes → those 12 agents chew through 180k tokens for $0.02 → Opus never reads raw agent output, only 3 summaries at 6k tokens the part most people get wrong: they let the expensive model do the reading. Haiku is 40x cheaper per token. mapping 48 files cost $0.009. doing that on Opus is where the whole budget goes. one catch worth knowing: past 100k of context Haiku jumps to 5x its own price, so every agent is kept under 100k on purpose. that is the whole design. the brain stays small and expensive, the hands stay huge and cheap, and nothing raw ever crosses between them. Opus plans. Haiku reads. Opus reviews. JEV engineering is the same move one layer down: 1,301 forks this run - which file, which agent, retry or stop - resolved in code at 0.90 sharpness, and only the split ones (0.48 on "which agent") ever reach Opus the full tree > Opus 5.5 runs the main session, plans the lanes and decides what happens next > 12 Haiku 5.5 agents fan out across read, grep and test, each kept under 100k context > every agent returns a summary, never raw files or logs > Opus 5.5 reads 3 summaries at 6k tokens and reviews the result > JEV resolves 1,301 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. Set every subagent to Haiku 5.5 and fan the grunt work out across read, grep and test lanes. 3. Keep each agent under 100k of context, split the work further if it goes over. 4. Agents return summaries only, raw files and logs never enter the main session. 5. Add one rule to CLAUDE.md: if a task is reading, searching or running tests, it never runs on Opus. Route deterministic forks (file choice, 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.

This post quotes another post: x.com/sprytixl/status/2107763347097534680