Claude Max API Credits: What $100 to $500 a Month Buys

On 7 October 2026 Anthropic started giving Claude Max 5x subscribers $100 a month in API credits, Max 20x subscribers $200, and Team plans up to $500 pooled, and unused credit disappears at the end of each billing cycle. At list prices, $200 buys roughly 667 runs of a 20K-input, 2K-output agent task on Fable 5.1, 3,333 on Sonnet 5.5, or about 66,666 on Haiku 5.5. The credits cover the API, Managed Agents and the Agent SDK. They do not cover interactive Claude Code. Below: the arithmetic per model, the exact coverage rules quoted from Anthropic, the Haiku 100K price cliff, three ways people burn the credit, and when Max 20x beats paying list price.

The change surfaced in the Haiku 5.5 launch discussion on Hacker News (1,047 points, 488 comments), and Simon Willison noted it the same day: Max 5x gets $100 a month, Max 20x gets $200, Team gets "up to $500, pooled across their users", and the credits "do not roll over".

What Anthropic actually promises

Anthropic's developer page, API credits for Max and Team plans, gives the amounts:

Plan	Monthly credit

| Max 5x | $100 |

| Max 20x | $200 |

| Team, Standard seat | $20 per seat, pooled |

| Team, Premium seat | $100 per seat, pooled |

Team credits for all seats pool into one balance capped at $500. Anthropic's own example: three Standard seats plus two Premium seats gives $260 a month. Free, Pro and Enterprise plans are not eligible. New subscribers can claim after 7 days on an eligible plan.

Claiming takes a few clicks. In claude.ai, open Settings, then Billing, and link a Claude Console organization in the API credits section. The catch: "Each plan links to one Console organization, and each Console organization can receive credits from one plan. You can't change the linked organization yourself." Pick the org you will actually build in. Credits then show up in the Console under Promotional credits, and no payment method is required on Claude Platform.

What draws on the credits and what does not

The developer page lists the covered products: the Claude API, Claude Managed Agents, the Claude Agent SDK and the playground. Not covered: Claude Code, extra usage in the Claude apps, and Claude on Bedrock, Google Cloud or Microsoft Foundry. It also says credits "don't change your usage limits in Claude or Claude Code."

That leaves the part everyone argues about: headless Claude Code. The Help Center article splits it by how you authenticate:

• With an API key from the linked Console org: "API credits cover claude -p and the Claude Agent SDK when you run them yourself with an API key."

• Signed in with your Claude plan: claude -p and Agent SDK usage "still draw from your plan's usage limits and don't use your API credits."

• Interactive Claude Code in the terminal, IDE, desktop or web: not covered.

• The Claude Code GitHub Action, IDE extensions and the desktop app: "count as Claude Code usage, so API credits don't cover them, even with -p."

One ambiguity. A second Help Center article says the credits "cover the Claude Agent SDK, claude -p, the Claude API, and Claude Managed Agents" and that you can still run those "with your subscription limits." Read together with the first article, the rule is the credential: API key from the linked org draws credits, plan sign-in draws plan limits. Anthropic does not state that rule in one sentence on either page, so test it. Run one cheap claude -p call with the key and watch the Promotional credits balance in the Console move.

If a request is not covered and the org holds only these credits, Claude Code shows "Credit balance too low." When the balance hits zero, "API requests stop until your next monthly API credits arrive" and usage is never charged to your plan. Every API key and workspace in the linked org draws from the same balance, so give each project its own workspace with a spend limit.

What $200 buys, model by model

Prices below come from Anthropic's pricing page, checked 13 October 2026. The reference task is a typical agent step: 20,000 input tokens (system prompt, tools, context) and 2,000 output tokens, no caching, no batch discount.

The arithmetic, once, for Fable 5.1 at $10 input and $50 output per million tokens: 20,000 / 1,000,000 x $10 = $0.20 for input, 2,000 / 1,000,000 x $50 = $0.10 for output, $0.30 per run. $200 / $0.30 = 666 full runs. Every other row is the same two multiplications.

Model	Input / output per MTok	Cost per 20K/2K run	Runs for $200	Runs for $100

| Fable 5.1 | $10 / $50 | $0.30 | 666 | 333 |

| Opus 5.5 | $4 / $20 | $0.12 | 1,666 | 833 |

| Sonnet 5.5 | $2 / $10 | $0.06 | 3,333 | 1,666 |

| Haiku 5.5 (prompt up to 100K) | $0.10 / $0.50 | $0.003 | 66,666 | 33,333 |

Read it as a budget in tokens instead and $200 buys 20M input tokens on Fable 5.1, 50M on Opus 5.5, 100M on Sonnet 5.5 and 2 billion on Haiku 5.5. On output, the same $200 buys 4M, 10M, 20M and 400M tokens.

The spread is the whole story. Fable 5.1 costs 100x what Haiku 5.5 costs per run on this task. Routing is worth more than any prompt trick. A pipeline that sends the planning step to Opus 5.5 and the 40 file-triage steps to Haiku 5.5 spends a fraction of an all-Opus run. Plug your own token counts into the AI Model Cost Calculator and model a multi-step loop in the Agent Run Cost Simulator.

The Haiku 100K price cliff

Haiku 5.5 is priced by prompt length. Up to 100,000 tokens it costs $0.10 / $0.50 per MTok. Over 100,000 it costs $0.50 / $2.50, five times more. Anthropic's pricing page says the prompt length "counts all of its input tokens, including cache reads and cache writes," and that each request is priced on its own, so a request over the threshold pays the higher price "even when part of its prompt is a cache hit."

Worked example, with a 2K-token reply:

Prompt size	Input cost	Output cost	Total

| 100,000 tokens | $0.0100 | $0.0010 | $0.0110 |

| 100,001 tokens | $0.0500 | $0.0050 | $0.0550 |

One extra token multiplies the bill by five. For a long-running agent that appends tool output every turn, the context eventually crosses 100K and the cheap model quietly stops being cheap. Trim or summarize history before it reaches the line. I covered the mechanics in the Haiku 5.5 pricing breakdown. Simon Willison also flags a second hit: the new tokenizer uses about 1.25x as many tokens as Haiku 4.5 on the same long prompt.

Three ways people waste the credit

1. Running everything on the top model

Defaulting to Fable 5.1 for classification, extraction and routing burns 100x the money Haiku 5.5 would. Use the expensive model for the one step that needs judgment.

2. Skipping caching and batch

Anthropic's credits page lists both as its own advice. Cache hits cost 0.05x base input on Opus 5.5 and Sonnet 5.5, and the Message Batches API "costs 50 percent less." A nightly eval suite that does not need an instant answer belongs in a batch. Remember the Haiku catch above: cache reads still count toward the 100K threshold.

3. Letting the balance expire

There is no rollover. A $200 credit used at 30 percent wastes $140. The opposite failure also happens: an unattended agent loop with no workspace spend limit drains the pool in a night, and every key in the org shares it. Set a workspace limit before you leave a loop running, and use up the credit on cheap-model batch work in the last week of the cycle.

A fourth, smaller trap: credits are consumed first, then purchased credits. Auto-reload watches only your purchased balance, so it can top up while included credit remains. Leave it off until you need it.

When Max 20x beats paying list price

The credit is a rebate on API spend you were going to make anyway, so the test is simple. If your monthly API bill at list price would exceed $200, Max 20x absorbs the first $200 of it. If the bill would be under $200, the credit is partly wasted and you should size down. I did not verify plan subscription prices on the pages above, so check the current price on claude.com/pricing before you compare.

Four profiles:

• You live in interactive Claude Code: the credits do nothing for that work. Your value is the plan limits, not the $200.

• You run scheduled agents or claude -p jobs with an API key: Max 20x credits are real money. At Sonnet 5.5 rates, $200 covers about 3,333 of the reference runs a month, over 100 a day.

• You run a Haiku 5.5 pipeline under 100K tokens per prompt: $100 already buys 33,333 runs, so Max 5x is more than enough and the extra $100 on Max 20x buys nothing you can use.

• You run Fable 5.1 on long contexts: 666 runs a month is about 22 a day. The credit disappears fast, and you will pay list price on top.

Teams have a different calculation. Premium seats carry $100 each, pooled to $500, so five Premium seats hit the cap. A sixth seat adds no credit.

If you want a ready-made sheet for tracking spend per model against the monthly balance, the Agent Cost Control Spreadsheet does the per-run and per-day math. For the Claude Code side of the bill, which these credits do not touch, see the Claude Code Production Pack and our Claude Code cost system guide.

To run these numbers for your own plan, model mix and token sizes, use the free Claude Max API Credits Planner. It shows runs per model, the day the credit runs out and whether the monthly reset arrives first.

Quick answers

Do the credits stack with my plan usage limits?

They are separate. Anthropic says credits "don't change your usage limits in Claude or Claude Code." Plan limits govern interactive use. Credits govern API-key usage on the Claude Platform.

Do unused credits roll over?

No. They expire at the end of each billing cycle, or monthly for annual plans. If you cancel or downgrade, credits you already hold stay usable until they expire, but new ones stop.

What happens when the balance reaches zero?

API requests stop until the next monthly credit arrives, unless the org has purchased credits or auto-reload on. Usage is never billed to your Claude plan.

Can I use the credits on Bedrock or Vertex?

No. Anthropic states the credits apply only to the Claude API in the Claude Console, not to Bedrock, Google Cloud or Microsoft Foundry.

Is the GitHub Action covered?

Anthropic says runs from the Claude Code GitHub Action count as Claude Code usage, so API credits do not cover them, even with -p. The page does not say in the same sentence which limit they draw on.

Pricing and coverage rules can change. Anthropic rolled this out over several days from 7 October 2026, so the claim button may not appear on every account yet. Re-check the two Anthropic pages linked above before you commit a budget to it.

Originally published at wowhow.cloud