并行跑 N 个 agent 的额度与成本：同一只池子，N 份上下文

不能这么指望，官方也没有这条承诺。到顶的行为是 Claude Code blocks further requests until the reset time shown in the message；官方给的续跑方式是 In an interactive session signed in with a claude.ai subscription, Claude Code can also wait in the open session and continue the interrupted task shortly after the reset —— 同一只订阅等重置，不是换号。换模型也不解，session and weekly limits are shared across all models。更要紧的是官方 Usage Policy 把 Create or manage multiple accounts to evade detection or circumvent platform safeguards 列为禁止项。周额度那层官方写 resets at a fixed time each week that is assigned to your account，与你何时开工无关（核于 2026-09-22）。

换模型这层官方有原句：Each model has its own prompt cache, so the next request re-reads the whole conversation with no cache hits。账号与组织这层，官方写的是 Caches are isolated between organizations. Different organizations never share caches, even if they use identical prompts —— 换到另一只组织，同一段前缀得重新写一次，而 5-minute cache write tokens are 1.25 times the base input tokens price。命中本身还要求 Cache hits require 100% identical prompt segments，分支名或每个 worktree 的路径混进前缀就不命中。至于社区流传的「打穿会涨 10 倍到 50 倍账单」，本次抓回的官方正文里核不到，属社区反馈、无官方口径，本页不当成本主张写（核于 2026-09-22）。

不一定，很可能更贵。每个 agent 都得自带一份固定前缀：官方写 Each subagent starts with a fresh, isolated context window，且 each teammate has its own context window and consumes tokens independently，token usage is roughly proportional to team size；workflow 那页另给方向——a single run can use meaningfully more tokens than working through the same task in conversation。按本页上面那组假设 P = 40,000，把一件事拆成 K 段就是把这份前缀付 K 次，这是本页算式的推论而非官方数字。官方唯一给过的倍数是 plan mode 下 agent teams 约 7x more tokens than standard sessions —— 那是 token 数，不是 7 倍账单金额。省与贵的判据落在同一处：省下的重复读入是否大过多付的 (K − 1) × P × 输入价（核于 2026-09-22）。

先说来源：80% 是编排器自己定的阈值，官方 usage-tracking 页原句 A warning chip when you cross 80% of a limit；Anthropic 官方示例里的预警文案是 85% of your session limit，两个数不是同一个东西。它盯的那一层是两条进度条：five-hour session and weekly usage limits，状态栏字段的定义是 Percentage of the 5-hour or 7-day rate limit consumed, from 0 to 100，并且这两个窗口 appears only for claude.ai Pro and Max subscribers，API key 侧不存在。读数本身也不实时：numbers update when the agent writes, not in real time。所以 80% 适合当刹车提示，不适合当对账依据（核于 2026-09-22）。

要分两只时钟。API 这侧：限额按每型号类的 RPM、ITPM、OTPM 计，超过回 429 并带 retry-after；官方另外两句正好打在并行上——Short bursts of requests can exceed the limit and trigger rate limit errors，以及用量陡增还会撞 acceleration limits。这只池子 Limits are set at the organization level，一次 fan-out 不会把容量变大，且它用 token bucket 持续回填而不是定点清零。订阅那侧到顶不是 429：Claude Code blocks further requests until the reset time shown in the message。而按量计费那一侧官方写 usage there is metered per request, so there is no reset to wait for —— 这一侧没有等重置的说法。本站 key 的并发是 5（截至 2026-09-22），排队不等于被封（核于 2026-09-22）。

照上面三步算例那一节做，再加三条限定。第一，单价必须带日期：官方价表页自己写着 Price and plans are subject to change at Anthropic's discretion，缓存那一档还要带 Prompt caching pricing reflects 5-minute TTL.。第二，命中要自己核：Cache hits require 100% identical prompt segments，前缀里有一个字节不一样就是没命中。第三，别拿编排器对账：它只覆盖 Claude Code、Codex、Gemini、OpenCode、Kimi Code、MiniMax 这 6 家，读的是各 agent 落在本地盘的记账（No API calls, no extra auth），估算价 not a live bill from the provider，官方自己让 Prefer the provider console for authoritative spend；同一账号的 claude.ai、Claude Code、Desktop 都 counts towards the same usage limit，所以要拼的是服务商或本站 key 的账，不是把几块状态栏加起来（核于 2026-09-22）。