Haiku 5.5 十万上下文价格悬崖：拆成 9 个小 agent 省 86%

Eliana 🇺🇸 @eliana_usa原推发布于 2026.10.27

在 X 上打开原推

Anthropic just shared one of the best tricks for Claude Code. The idea: Haiku 5.5 has a price cliff at 100k context. Cross it and every token costs 5x more. So you never let an agent get there. Same work, two ways: one big agent would have cost $6.15. Nine small ones came out at $0.85. That is 86% saved on the exact same job. You just launch Claude Code like this: claude --model opus # agents: model: haiku, each < 100k here's the split: → Opus 5.5 leads, splits the work into lanes and merges at the end → Haiku 5.5 agents do everything else, each one kept under 100k → when an agent nears the cliff it forks in two instead of growing → Opus reads 9 summaries, never raw agent context the numbers that matter: under 100k Haiku is $0.10 in and $0.50 out per million. over 100k it jumps to $0.50 and $2.50. same model, same tokens, 5x the bill, and nothing warns you when it happens. in this run a4 was at 90k with 10k left and a7 at 84k with 16k left. both got forked before they crossed. a11 started fresh at 41k on the cheap tier. 9 agents, 6 splits, and exactly 1 crossed the line anyway. that one agent is the reason the bill was $0.85 instead of $0.40. cache reads count toward the 100k too, which is what catches most people out. Opus leads. Haiku swarms. Nobody crosses the cliff. JEV engineering is the same move one layer down: 1,900 forks this run - split or keep, which agent, retry or stop - resolved in code at 0.93 sharpness, and the split-or-keep call at 0.42 is exactly where a fork gets triggered the full tree ▸ Opus 5.5 leads the session, splits the work and merges at the end ▸ every subagent runs on Haiku 5.5 with a hard 100k context ceiling ▸ an agent nearing 100k forks in two instead of growing past it ▸ each fork re-reads only the context it actually needs ▸ Opus 5.5 reads summaries only, never raw agent context ▸ JEV resolves 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 as lead: split work, merge results, write no code. 2. Set every subagent to Haiku 5.5 with a hard context ceiling of 100k, cache reads included. 3. When an agent approaches 100k, fork it into two agents instead of letting it grow. 4. Each fork re-reads only the context it needs, never the full history. 5. Add one rule to CLAUDE.md: no agent is ever allowed to cross 100k, no exceptions. 6. Route deterministic forks (split or keep, agent choice, retries) through JEV. Show me every change as a diff first. No edits until I say go." ↳ https://code.claude.com/docs/en/model-config save it, this one line is worth 86% of your bill.