Claude Max API Credits: What $200 a Month Really Runs

Anthropic added the credits on October 7, the same day Haiku 5.5 shipped, and it got a lot less attention than the model. It should get more. If you already pay for Max to use Claude Code, you now get a monthly API budget you didn't have to buy separately. Here's the breakdown from the official credits page:

A team with three Standard and two Premium seats gets $260 a month. Plans bought through the iOS or Android app count too, but you claim them on claude.ai in a browser.

The credits cover four things: the Claude API, Claude Managed Agents, the Claude Agent SDK and the Console playground. They don't cover Claude Code, extra usage inside the Claude apps, or Claude running on Bedrock, Google Cloud, Microsoft Foundry or Claude Platform on AWS. They also don't change your Claude or Claude Code usage limits. It's a separate wallet, not a bigger one.

"$200 of API" means nothing until you put a workload next to it. So I used a real one: the agent that researches, writes, builds and publishes posts on this blog. I pulled the token counts for one complete run from my agent's session log. It made 90 model calls and 108 tool calls over about two hours, and it used:

• 261,706 new input tokens
• 11,597,312 cache-read tokens (the conversation history it re-reads every turn)
• 28,994 output tokens

Then I priced that run at the published Claude API rates, with new input billed as 5-minute cache writes because that's how an agent loop with caching behaves:

A daily agent on Opus 5.5 fits inside Max 20x with about $74 to spare. The same agent on Fable 5.1 doesn't fit at all. On Max 5x, Sonnet 5.5 is the realistic ceiling for anything that runs every day.

Two things in that math surprised me. First, cache reads were 55% of the bill on both Sonnet 5.5 and Opus 5.5. Long agent loops re-read their history on every call, and 11.6 million cached tokens add up even at a twentieth of the input price. That's also why the October 7 cut to Sonnet 5.5 cache reads ($0.20 down to $0.10 per million) matters more than it sounds: it took this run from $3.26 to $2.10. I wrote more about how the cache math works in my prompt caching breakdown.

Second, Haiku 5.5 doesn't get its headline price on a workload like this. The average call carried about 132,000 tokens of context, which is over the 100K line where Haiku 5.5 jumps to $0.50 input and $2.50 output for the whole request. Priced on the high tier it's $0.82 a run; the under-100K tier would have been $0.16. If you're planning to stretch credits with Haiku, trim the context first. I measured the tokenizer side of that in Haiku 5.5 vs Haiku 4.5.

No, and this is the line most people will trip on. The Agent SDK is covered, Claude Code isn't, even though one is built on the other. If you write a TypeScript or Python agent on the Claude Agent SDK and authenticate with an API key from the linked org, the credits pay for it. If you point Claude Code at that same key, it shows Credit balance too low · Add funds once the org has nothing but these credits.

The practical rule: don't export ANTHROPIC_API_KEYin your shell profile, or Claude Code can end up using it instead of your Max login. Load the key per project, from that project's .env, only in the processes that are supposed to spend API credits. If you're still deciding between the SDK and plain API calls, I compared them in Claude Agent SDK vs the API.

It takes a couple of minutes, but one choice in the middle is permanent.

1. On claude.ai, open Settings > Billing (Max) or Organization settings > Billing (Team).
2. In the API credits section, pick the Claude Console organization that should get the money and accept the program terms.
3. In the Console, check Settings > Billing for a Promotional credits line with the amount and expiry date. Then create a key.

Each plan links to one Console org, each org takes credits from one plan, and you can't change the link yourself. Only support can move it. If your current org is shared with a client, a contractor or an old side project, make a fresh one before you click. You need to be the Max subscriber (or a Team Owner) and Owner, Admin or Billing on the Console side. New subscribers wait 7 days before the option shows up.

1. "Ran out" and "never funded" look the same. When the balance is empty the API returns Your credit balance is too low to access the Anthropic API. The key my automations use returned that exact string when I tested it this morning. From the response alone you can't tell whether the included credits ran out, were never claimed, or went to a different org. Check the Promotional credits line in the Console before you debug anything else.

2. Auto-reload ignores your included credits. Included credits are spent first, but auto-reload only looks at the balance you bought. If that purchased balance is under your reload threshold, your card gets charged even while $180 of free credit is sitting there. Lower the threshold or turn auto-reload off if the credits are meant to be the budget.

3. Two different month boundaries.Spend caps reset on the 1st of the calendar month. Credits refresh on your billing date. If you bill on the 20th, a cap you set around the credit amount won't line up with it.

4. One balance for every key. Every workspace and key in the linked org draws from the same pool. One runaway retry loop can eat the month for everything else. Give each app its own workspace with a spend limit. I go through what that looks like on a cron agent in cost controls for autopilot agents.

5. The tier bump is smaller than it looks. A linked org moves to at least the Start tier, and to Build only if its monthly credits are over$200. Read literally, Max 20x at exactly $200 doesn't clear that bar, and a Team pool of $260 does. The credits also don't count toward earning a higher tier, so if you hit rate limits, see how I handle Claude 429s.

My take: if you already pay $200 for Max 20x, this is $2,400 a year of API spend you were going to buy anyway. Use all of it. Credits that expire unused are the only real waste here.

• Sonnet 5.5 for agent loops.It's the best cost per run now that its cache reads are $0.10.
• Haiku 5.5 at effort low for the cheap plumbing, like classification, extraction and routing, with prompts kept well under 100K.
• Opus 5.5 only where Sonnet failed a real test. Fable 5.1 on a credit budget is a once-in-a-while tool, not a cron.
• Batch the backlog near the end of the cycle. The Batch API halves the price, and evals, re-summaries and enrichment jobs can wait a day. Whatever credit is left the week before renewal goes there.

The token counts come from one real run of my blog agent on September 1, 2026, read from the agent's session database: 90 API calls, 261,706 uncached input tokens, 11,597,312 cache-read tokens and 28,994 output tokens (7,495 of them reasoning). That run used a non-Claude model, so treat these as workload-shaped estimates, not a Claude invoice. Claude's tokenizers will count the same text differently, and Haiku 5.5 counts noticeably more. Prices are the published Claude API rates on October 11, 2026, with uncached input billed as 5-minute cache writes. The Haiku 5.5 row uses the over-100K tier because the average call carried about 132K tokens. US-only inference adds 1.1x to everything and isn't included.