06-reference/research

headless cron vs managed agents metering

2026-10-19·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
claude-managed-agentsmeteringmax-planapi-creditscost

Separate pools, by auth method: the Mac crons spend subscription quota, a Managed Agents pilot spends an API credit the crons cannot touch

The question

"Does a headless claude -p run on the Mac mini cron (studio rounds, nightly skills) draw on the same Max-plan $200/mo credit pool as a Claude Managed Agents pilot, or are they billed/metered separately?"

This is the fact flagged "confirm before pilot step 3" in [[2026-10-12-managed-agents-liftability-assessment]]. It gates whether a Claude Managed Agents (CMA) pilot is additive spend or competes with the existing cron workload.

What we already know (from the vault)

What the web says

All four pages below were read 2026-10-19. Where a page shows only a relative date, that is stated.

Documented by Anthropic. "Monthly API credits for Max and Team plans," Claude Help Center, https://support.claude.com/en/articles/17154008 (shows "Updated over a week ago," no exact date):

Documented by Anthropic. "Use Claude Code with your Pro or Max plan," https://support.claude.com/en/articles/11145838-use-claude-code-with-your-pro-or-max-plan ("Updated this week"):

Documented by Anthropic. "Use the Claude Agent SDK with your Claude plan," https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan: now carries an October 7 2026 notice above the June 15 2026 pause notice. The October notice says Max and Team plans "now include monthly API credits, which cover the Claude Agent SDK, claude -p, the Claude API, and Claude Managed Agents." The June pause text survives below it and still says "You can still use the Claude Agent SDK, claude -p, and third-party apps with your subscription limits." Read together with article 17154008, the two are not in conflict: the credit covers claude -p under API-key auth, the subscription covers it under plan login.

Documented by Anthropic (bundled claude-api skill, shared/managed-agents-core.md, read 2026-10-19). A CMA session is addressed inside a Console workspace that an API key belongs to, which is the same billing surface the credits land in. CMA list cost counts three things, not just tokens: model tokens at each model's list price, web searches at $10 per 1,000, and session running time at $0.08/hour. A session budget is a hard ceiling priced at public list rates; at the cap the session pauses idle with stop_reason: budget_reached rather than terminating, and raising or removing the budget resumes it.

Third-party, not load-bearing. Secondary coverage (mixed-news.com, explainx.ai, qcode.cc, ssdnodes.com, intuitionlabs.ai) dates the credit launch to October 7 2026, reports that the Claude plan is never billed for API overage (the API simply stops when credits run out), and reports that Pro, Free and Enterprise are not eligible. The overage-stop behavior is the one operationally important item here that I could not confirm in a first-party page, so treat it as unconfirmed. Third-party per-window message counts and rolling-versus-fixed weekly reset claims conflict with each other and with the help pages; ignore them.

Where the sources are silent. No first-party page states a concurrency limit for subscription Claude Code sessions, a concurrency limit for CMA sessions, or whether cron-driven headless use is affirmatively permitted under the subscription. The Consumer Terms automation clause remains the only text on point, and it grants no explicit permission.

Convergences and contradictions

Synthesis for RDCO

They are metered separately, and the separation is cleaner than the pilot plan assumed. An interactive Claude Code session on Max draws the plan's usage limits. A headless claude -p run draws those same plan limits when the CLI is signed in with the Claude plan, which is how every Mac mini cron runs today, and draws monthly API credits only when it is run with an API key from the linked Console organization. CMA has no subscription-login path at all: it is addressed through a Console workspace, so it always draws the credit, then purchased credits, and never the plan quota. The founder's two $200 numbers are linked but independent: paying for Max 20x is what makes the credit grant exist, and the grant lands in a different ledger that the subscription side is forbidden from touching.

So a CMA pilot is additive, not cannibalizing - with one condition and one inversion. The condition: the Console organization has to actually be linked, because the grant is claimed manually and only once, and the link cannot be changed later without support. The inversion worth naming: because the credit does not roll over and expires at the end of each billing cycle, a Max 20x subscriber who has never linked an organization and never run an API-key workload is forfeiting the whole grant every month. On that reading the pilot is not new spend at all - it is the first use of an allowance already bought. The nightly-deep-research shape costed in [[2026-10-12-managed-agents-liftability-assessment]] was sized to fit inside the grant, and nothing in that sizing has to compete with the crons. Section 5's "what would make this wrong" clause can be closed: the Mac crons do not draw on the credit.

The thing that actually constrains RDCO is the subscription quota, not dollars, and the pilot relieves it rather than straining it. The binding ceiling on the Mac is the five-hour window plus the weekly cap, shared across every Ray session and the founder's own chat and terminal use. Moving a nightly job to CMA frees window and weekly headroom on the Mac and spends a dollar balance that was otherwise going to expire. The reverse move - switching the crons to API-key auth to "use the credit" - is the only way to create the competition the backlog note feared, and it would also surrender the subscription quota the crons currently run on for free. Do not do that as a side effect of setting up the pilot: keep the cron login and the CMA key strictly separate.

One finding outranks the cost math, and it points the opposite way from the usual caution. Per [[2026-09-13-anthropic-aup-agentic-use-operative-clauses]], the Consumer Terms permit automated access only via an API key or where Anthropic "otherwise explicitly permit[s] it," and no operative written permission for cron-driven claude -p on a plan login has been found at the source. The first-party pages read today do not close that gap: article 11145838 is silent on headless and automated use, and article 17154008 only describes how plan-login claude -p is metered, which is acknowledgement rather than permission. A CMA pilot runs on an API key under the Commercial Terms, which is the side of that clause with an explicit carve-out. The pilot therefore reduces terms exposure for whatever work moves onto it. The practical caps to respect on the CMA side are different in kind: a session budget is a hard list-cost ceiling that pauses a session mid-work (a run cut off at the cap is a failed fire, as section 5 already says), and CMA bills session running time at $0.08/hour and web search at $10 per 1,000 on top of tokens, so a long idle-but-alive session spends the grant even when the model is quiet.

Why this is in the vault

This closes the single gating fact on pilot step 3 of [[2026-10-12-managed-agents-liftability-assessment]] - whether a CMA deep-research deployment competes with the Mac mini cron workload - and it retires the incorrect separate-credit claim in [[2026-07-04-anthropic-max-plan-tos-productized-agentic-use]] that caused the uncertainty. It also changes the pilot's framing from "approve new spend" to "claim an allowance that currently expires unused," which is a different decision for the founder.

Open follow-ups

Test plan: the exact check that settles what is still open

Two things remain account-specific rather than documented, and both are read-only.

  1. Is the Console organization linked, and which one? On claude.ai: Settings > Billing > the "API credits" section. It shows either a "Link organization" button (not claimed) or the linked organization name (claimed). Requires Owner, Admin or Billing role in that Console organization, and the same email as claude.ai. This is founder-side; the link is one-time and not self-changeable, so pick the organization the pilot will live in before confirming.

  2. Has anything already consumed this cycle's grant? On the Console: https://platform.claude.com/settings/usage filtered to the current billing cycle, cross-checked against https://platform.claude.com/settings/billing for the credit balance. A zero-consumption cycle confirms the crons are not touching it.

  3. Confirm the cron auth path from the machine, not from notes. Runnable verbatim on the Mac mini:

grep -rn -E 'ANTHROPIC_API_KEY|ANTHROPIC_AUTH_TOKEN' ~/.claude/scripts/cron-job-runner.sh ~/.claude/scripts/cron-dispatcher.py; ant auth status

Expected: no match in either file, and ant auth status reporting an active OAuth profile or no API-key credential source. A match means a cron is already billing the credit and the additive conclusion above does not hold for that job.

What proceeded without the answer. The metering topology question is now answered from first-party pages, so pilot steps 1 and 2 of [[2026-10-12-managed-agents-liftability-assessment]] (scoping the allowlisted qmd collection, defining the agent) need no further gating and can proceed. Only step 3, the scheduled deployment, depends on check 1 above.

Cost of waiting. Each unclaimed billing cycle forfeits that cycle's entire grant, because credits do not roll over and expire at cycle end. With renewal around the 15th, waiting a month costs one full grant and buys nothing, since the check is a read-only look at two settings pages.

Related

Sources