Claude API Credits on Max: What $200 Covers, and When claude -p Counts

On October 7, Anthropic added Claude API credits to the Max and Team plans. Max 5x gets $100 a month, Max 20x gets $200, and a Team plan gets up to $500 pooled.

I run my agents headless, so my first question was simple. Do my claude -p workers draw from that credit now?

Only if the process holds an API key from the Console organization you linked. Signed in with your plan, nothing changes. The -p flag doesn't decide who pays. The credential does.

What the Claude API credits cover

Anthropic's platform docs put it in one table:

The credit works on any model. On Team it is $20 per Standard seat and $100 per Premium seat, pooled into one balance and capped at $500.

This is not the Agent SDK credit Anthropic announced in May and paused in June. I built a standby lane for that one, described in the CLI subprocess pattern that survived the OAuth ban. The help center answers that directly: "No. That credit isn't available."

Your plan limits don't move either. The docs say the credits "don't change your usage limits in Claude or Claude Code."

How to claim them and check your balance

The credits don't appear by themselves. You link a Claude Console organization to your plan, and the credits land there.

1. On claude.ai, open Settings > Billingon Max, orOrganization settings > Billing on Team.
2. In the API credits section, pick the Console organization that should receive them.
3. In the Console, open Settings > Billing. The balance sits under Promotional credits, with the amount and the expiry date.

Three things to know before you click.

Pick the organization on purpose. The docs warn: "You can't change the linked organization yourself, so link the organization you plan to build in." Changing it later goes through support.

New subscribers wait. You can claim "after 7 days on an eligible plan."

The section may not be there yet. Anthropic's help article says: "We're rolling out API credits over a few days so we might not have gotten to you yet."

You don't need a payment method on the Console to use them.

Does claude -p use the credits?

This is where most of the coverage goes wrong, in both directions. The help article has the rule:

"API credits cover claude -p and the Claude Agent SDK when you run them yourself with an API key from your linked Claude Console organization, because that usage is billed as Agent SDK usage."

And the other half:

"When you're signed in with your Claude plan instead, claude -p and Agent SDK usage still draw from your plan's usage limits and don't use your API credits."

So the same command can land in two different wallets:

My own orchestration repo used to carry the wrong version of this: "claude -p is headless, so it is API-metered." I took that sentence out in August, after checking what a worker was authenticated as. There was no key in its environment and a plan login in the keychain. The run was on the subscription.

You can check your own machine in one command:

The JSON has an authMethod field. claude.ai means the plan login is in use. api_key means a key is winning. On my machine today it says claude.ai with subscriptionType: max, and there is no ANTHROPIC_API_KEY in the shell. Every headless worker I run is on the plan.

The trap: a Claude Code API key in your environment beats your plan login

Claude Code picks one credential from a fixed order. ANTHROPIC_API_KEY is third. Your plan login is seventh. The authentication docs add one sentence that matters for anyone who scripts:

"In non-interactive mode (-p), the key is always used when present."

Interactive sessions ask you once whether to use the key. Headless runs don't ask.

Approve it in an interactive session on an organization that only holds these credits, and Claude Code stops with "Credit balance too low". The credits don't cover Claude Code.

That cuts both ways.

A key in a shell profile moves everything. Export it in .zshrc or a shared env file and every headless run in that environment bills the Console organization. Child processes inherit it too. There is no error and no notice.

Then the balance runs out. The docs describe what happens: "API requests stop until your next monthly API credits arrive." Claude Code shows Credit balance is too low. A build worker that hits that halfway through is lost work.

Unless auto-reload is on, and then you pay. Included credits are spent first. But auto-reload looks only at your purchased balance: it "reloads when that balance reaches your threshold, even if included credits remain."

And the opposite mistake costs you differently. If you think a job is on credits but the process is signed in with the plan, it eats your plan limits. In September one of my headless workers ended with "You've hit your session limit · resets 1pm". From the outside it looked like a worker that forgot to write its report. Credits don't fix that unless the key is set.

What I do about it:

• The key lives in the command of one process. Not in a profile, not in an env file that other jobs read.
• A separate config directory per identity. With CLAUDE_CONFIG_DIR set, Claude Code keeps its own credential entry for that directory. I run two accounts on one Mac that way, written up in Dutch in Twee Claude-accounts op één Mac.
• A workspace with a spend limit. Every key in the linked organization draws from the same balance. The docs suggest a workspace per project with its own limit.
• Check the balance, not the endpoint. On another provider, my availability probe said the API was fine. The job died eleven seconds later on an empty balance.

A single job on the credits then looks like this:

That example reads the key from the macOS keychain. Use whatever secret store you have. The total_cost_usd in the output is an estimate Claude Code computes on your side, not a bill.

What $200 buys in a real orchestration

I looked up what my own Claude work would cost at API list prices. My orchestration writes a receipt per dispatch, and part of those carry a price that Claude Code computed from token counts.

These are list-price equivalents for work that ran on my subscription. I did not pay these amounts.

Together that is 211 dispatches and $254 in 23 days. And that is only the priced part. Of the 4,370 receipts written since mid-September, 545 carry a price at all, across every provider.

So $200 is roughly 320 Sonnet 5.5 dispatches of my average size, or about 50 Opus ones. For build work the credit does not replace the subscription. It is a second, smaller budget.

One limit on these numbers. On two non-Anthropic lanes I compared the receipt estimate with real spend, and it was off by a factor of 5 to 18. Cache hits and a proxy were miscounted there. For Claude I have not reconciled it against a bill, so read the table as an order of magnitude.

Where I'd spend it

Not on builders. A credit that runs dry in the middle of a build costs more than it saves.

I would start with jobs that are small per call and fail open: tagging against a fixed vocabulary, a scout pass before a dispatch, a nightly summary. If the balance is gone, the main work continues without them.

Claude Haiku 5.5 makes that budget go far. It costs $0.10 per million input tokens and $0.50 per million output tokens. This week I ran 450 Dutch business texts through it for a classification benchmark. The whole run cost $0.10.

One thing to watch there. Haiku 5.5 is priced by prompt length: above 100,000 tokens a request pays five times more. My review seats accept diffs of up to 800,000 characters, roughly 200,000 tokens. A Haiku 5.5 reviewer on a large diff crosses that line and pays the high rate.

I have not moved a lane yet, and Haiku 5.5 is not in my model registry. When I do, it gets its own lane with its own billing label, so a receipt shows which wallet paid. I keep one hard rule in my setup: the line runs along the auth identity, not along the package. The plan login and the Console key are two identities, and they never share a process.

If you use a third-party tool with your Claude login instead, different rules apply. I wrote those up in Hermes Agent Can Use Your Claude Subscription. That Doesn't Make It Allowed.