Monthly API credits for Max and Team: the same $200, the opposite sign

On 7 October 2026 Anthropic started handing Max and Team subscribers a monthly balance of Claude Platform credits. $100 a month on Max 5x, $200 on Max 20x, and on Team $20 per Standard seat plus $100 per Premium seat, pooled into one balance capped at $500. You link a Console organisation to your plan, accept the credit terms, and the money appears under Promotional credits. No payment method required.

At annual seat rates the credit is exactly the price of the subscription, which is the kind of number that makes you go back and read the terms. The terms are worth reading, because five months ago these same two figures were pointing the other way.

The same numbers, from the other direction

On 13 May 2026 Anthropic announced that programmatic usage — claude -p, the Agent SDK, third-party apps — would move off subscription usage pools and onto exactly this ladder of monthly credits, with API billing beyond it. Reported amounts varied at the lower tiers, but the top two matched what has now shipped: $100 on Max 5x, $200 on Max 20x. It was a ceiling. On 15 June, the day it was due to take effect, Anthropic paused it, and the support page went back to saying that for now nothing had changed.

October's version keeps the amounts and inverts the sign. Programmatic usage signed in with your plan still draws on your plan's limits, untouched. The credits are a separate API balance on top, and the docs say so twice: credits don't change your usage limits in Claude or Claude Code, and usage is never charged to your Claude plan. What was going to be a cap on your subscription is now a grant beside it.

That makes it additive, and genuinely good. It also means everything interesting is in the small print, because an additive grant is defined by what it refuses to pay for.

What the grant is worth

Pay annually and Anthropic is returning the full sticker price of the seat as API credit; pay monthly and it is four fifths. Discounted Team plans for nonprofits and scientists get the same per-seat amounts. Anthropic's own worked example is three Standard and two Premium seats, which comes to $260 a month, and adding one more Premium seat takes the next cycle to $360 — each month's credit is computed from the seats on the plan at the start of that billing month, so a seat added today pays out next month.

the Team ceiling, whatever the headcountreached at 25 Standard seats

$500

of the annual seat price, back as credit80% if you are billed monthly

100%

credit you must exceed for Build tierMax 20x lands exactly on it

$200

ever billed to your Claude planrequests stop instead

$0

The question everyone asks first

Does it pay for Claude Code? No — and the reason is worth understanding, because it is not a product name on a blocklist. The coverage table in the docs reads: Claude API yes, Managed Agents yes, Agent SDK yes, Playground yes, Claude Code no, extra usage in the Claude apps no, Bedrock and Vertex and Foundry no.

Which looks self-contradictory, because the Agent SDK is Claude Code packaged as a library and claude -p is Claude Code with a flag. The Help Center supplies the missing rule: what decides is the billing path, not the binary. A run authenticated with an API key from the linked Console organisation is API usage and draws on the credit. The same run signed in with your Claude plan draws on your plan's limits. And a run started by the Claude Code GitHub Action, an IDE extension or the desktop app counts as Claude Code either way, -p or not.

How far $200 actually goes

Far enough to build; not obviously far enough to run. The variable that decides which is not the model you pick.

Cache reads are $0.20 per MTok on both Opus 5.5 and Sonnet 5.5, so on a session-shaped workload the prefix costs the same on either model and the tier choice barely moves the count. Forget cache_control and you are paying $4 or $2 per MTok to re-read the same context every single turn, and the month's grant is gone in a week. That is a one-line fix worth roughly nine times what agonising over the model tier is worth.

// $ per MTok on the Claude API. The cache-read row is the one that
// decides how long a monthly grant lasts.
const PRICE = {
  "claude-opus-5-5":   { input: 4, cacheRead: 0.2, output: 20 },
  "claude-sonnet-5-5": { input: 2, cacheRead: 0.2, output: 10 },
} as const;

/** Included credit per month. Team pools $20 a Standard seat, $100 a Premium, capped at $500. */
const GRANT = { "max-5x": 100, "max-20x": 200 } as const;

type Turn = { cached: number; fresh: number; out: number };

function turnsPerMonth(
  model: keyof typeof PRICE,
  plan: keyof typeof GRANT,
  t: Turn,
) {
  const p = PRICE[model];
  const perTurn =
    (t.cached * p.cacheRead + t.fresh * p.input + t.out * p.output) / 1_000_000;
  return Math.floor(GRANT[plan] / perTurn);
}

// A mid-session agent turn: 200K cached prefix, 4K fresh, 2K out.
turnsPerMonth("claude-opus-5-5", "max-20x", { cached: 200_000, fresh: 4_000, out: 2_000 });
// 2083. Move that same 200K from cached to fresh and it is 233.
// Run it against your own usage logs before deciding the grant is or isn't enough.

Haiku 5.5 is the other lever, and a bigger one than the chart suggests. At $0.10 in and $0.50 out, a 10K-token prompt with a short answer costs a fifth of a penny, so $200 is somewhere around a hundred thousand classification, routing, tagging or extraction calls a month. For that half of a workload this is not a trial allowance. It is a production budget, and the Batch API halves it again on anything that tolerates waiting.

Four lines in the docs that decide what it is worth

• Two calendars. Credits refresh on your billing cycle. Your organisation's monthly spend cap resets on the first of the calendar month, and credit-funded usage counts against it. If your plan renews on the 20th, the fresh grant lands inside a cap that is already two thirds spent.
• Over $200 is strict. Linking an eligible plan lifts the organisation to at least the Start rate-limit tier, and a credit over $200 a month lifts it to at least Build. Max 20x grants exactly $200, so it does not clear the bar; Anthropic's own five-seat example team, at $260, does. Neither amount counts toward climbing a tier on spend.
• One organisation, one plan, for good. Each plan links to one Console organisation and each organisation can receive credits from one plan. You cannot change it yourself — that is a support ticket. Five Max plans cannot be pooled into the company org, and linking a personal plan to the production org is a decision you make once.
• Auto-reload watches the wrong balance. It triggers on your purchased balance only, so it can buy credits while $180 of included grant sits unspent. It is also the only thing standing between an empty grant and a stopped application.

The Team pool is quietly regressive

$20 a seat is generous on a small team, and the ceiling arrives sooner than it looks. Twenty-five Standard seats hit $500. Every seat after that is credited at nothing.

Which is coherent once you read the programme as what it plausibly is — a funnel. A Max subscriber with a linked organisation, an API key and a running app is a Platform customer, and the teams whose apps outgrow $500 a month are precisely the ones Anthropic wants on purchased credits. Excluding interactive Claude Code is the same logic from the other end: it stops the grant becoming a cheaper way to buy Claude Code capacity.

What to actually do

1. 1Link the organisation you will build in, not the one you happen to be signed into. It is one-way without a support ticket.
2. 2Give every experiment its own workspace with a spend limit. Every key in the linked organisation draws on the same balance, so one runaway loop in a prototype spends the month on everyone's behalf.
3. 3Turn on prompt caching before you tune anything else. On the credit's lifespan it is worth about nine times what the model choice is worth.
4. 4Put a purchased balance or auto-reload behind anything with users. The grant's failure mode is a stopped application, not a surprise invoice.
5. 5Batch what can wait and route what can be routed. Half price plus Haiku 5.5 is what turns a build budget into a production one.
6. 6Spend it or lose it. Nothing rolls over, so a quiet month is not saved up for a busy one.
7. 7Write down both dates — the billing-cycle day your credits refresh, and the first of the month when the spend cap resets. They are not the same day, and only one of them is on your renewal notice.

In May these numbers were a ceiling on what you were allowed to run. In October they are a budget for what you build next. The figures did not move; the direction did.

None of this is a discount. If you never call the API the credit is worth nothing and your subscription costs what it cost last month. What changed is that the price of starting a real Platform project dropped to zero for anyone already paying for Max — no card, no procurement, no approval — and the grant is large enough to carry a genuine production workload on the cheap models and a couple of thousand agent turns on the expensive ones.

Claim it, link the right organisation, then spend the first hour switching on prompt caching. That single setting decides whether the grant lasts the month or the week.