Parallel agents quota and cost: one pool, N contexts

Not as a promise. What the docs specify is Claude Code blocks further requests until the reset time shown in the message, and the sanctioned way to resume is 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 — the same subscription waits out the reset, it does not hop accounts. Swapping models does not help either: session and weekly limits are shared across all models. More importantly the official usage policy lists Create or manage multiple accounts to evade detection or circumvent platform safeguards as prohibited. The weekly layer, per the docs, resets at a fixed time each week that is assigned to your account, independent of when you started (checked 2026-09-22).

For models the wording is explicit: Each model has its own prompt cache, so the next request re-reads the whole conversation with no cache hits. For accounts, what the docs actually say is Caches are isolated between organizations. Different organizations never share caches, even if they use identical prompts — move to another organization and the same prefix is written again, at 5-minute cache write tokens are 1.25 times the base input tokens price. Even staying put requires Cache hits require 100% identical prompt segments, so a branch name or a per-worktree path inside the prefix defeats the hit. The community figure of a 10 to 50 times bill has no official counterpart in the pages captured here: it is community feedback without an official wording, and this page does not write it as a cost claim (checked 2026-09-22).

Not necessarily; often the opposite. Every agent carries its own copy of the fixed prefix: Each subagent starts with a fresh, isolated context window, and each teammate has its own context window and consumes tokens independently, with token usage roughly proportional to team size; the workflows page adds that a single run can use meaningfully more tokens than working through the same task in conversation. Under this page's assumption P = 40,000, splitting one job into K parts pays that prefix K times — our arithmetic, not an official figure. The only multiplier Anthropic itself gives is plan mode agent teams at 7x more tokens than standard sessions, and that is tokens, not a bill seven times larger. The break-even test is one line: whether the duplicate reads you avoid exceed (K − 1) × P × the input price (checked 2026-09-22).

Start with provenance: 80 percent is the orchestrator's own threshold — its usage-tracking page says A warning chip when you cross 80% of a limit — while Anthropic's own example wording is 85% of your session limit; the two numbers are not the same thing. The layer it watches is a pair, five-hour session and weekly usage limits, defined mechanically as Percentage of the 5-hour or 7-day rate limit consumed, from 0 to 100, and those windows appears only for claude.ai Pro and Max subscribers, so an API key has neither. The readout itself is not live: numbers update when the agent writes, not in real time. Fine as a brake light, wrong as a reconciliation basis (checked 2026-09-22).

Two clocks again. On the API side limits are measured per model class in RPM, ITPM and OTPM, and going past them returns 429 with a retry-after; two official sentences land straight on parallelism: Short bursts of requests can exceed the limit and trigger rate limit errors, and a sharp jump in usage can also hit acceleration limits. That pool is sized where Limits are set at the organization level, so a fan-out never enlarges it, and capacity is refilled continuously rather than being reset at fixed intervals. On the subscription side the cap is not a 429: Claude Code blocks further requests until the reset time shown in the message. For API keys, cloud hosting and pay-as-you-go the docs say usage there is metered per request, so there is no reset to wait for — there is nothing to wait out. Our own key allows concurrency 5 (as of 2026-09-22); queueing is not a ban (checked 2026-09-22).

Follow the three-step example above, then add three limits. First, dates on unit prices: the official pricing page itself states Price and plans are subject to change at Anthropic's discretion, and the cache tier carries Prompt caching pricing reflects 5-minute TTL. Second, verify hits yourself: Cache hits require 100% identical prompt segments, so one differing byte in the prefix means no hit. Third, do not reconcile against the orchestrator: it covers 6 providers (Claude Code, Codex, Gemini, OpenCode, Kimi Code, MiniMax), it reads bookkeeping each agent writes to local disk (No API calls, no extra auth), its cost line is not a live bill from the provider, and the docs say Prefer the provider console for authoritative spend. On one account claude.ai, Claude Code and Desktop share one allowance, so you assemble the provider or key bill, not a sum of status bars (checked 2026-09-22).