The Anthropic API bills on a pre-paid credit model. You add money to your account balance, requests draw from that balance in real time as they run, and when the balance runs out the API stops responding until you top up. That's the whole model — no monthly minimum, no seat fees, no per-model subscription. This piece covers how credits actually behave in practice and the four things that surprise most new accounts.
What a credit is
A credit is one U.S. dollar of API balance. When you top up $50, your account balance shows $50 in credits. When a request costs $0.023, your balance drops by $0.023. There's no unit conversion, no expiring tokens, no separate credits for different models. It's a running account.
Adding credits
- In Settings → Billing, click Add Credit. Minimum top-up is $5. Maximum per single top-up is $500 for most accounts, higher for verified organizations.
- Payment methods: credit card, U.S. debit card, and (for organizations) invoice billing after a credit check.
- Credits do not expire. Money on the account stays there indefinitely — including across gaps in usage.
Auto-recharge
Turn this on for anything running unattended. Set a threshold ($10 is the common default) and a recharge amount ($50–$200). When your balance drops below the threshold, Anthropic charges your card for the recharge amount. Without auto-recharge, an active workflow can burn through a balance overnight and the next request errors with `insufficient_quota` — the workflow just stops until a human notices.
Usage tiers
Every Anthropic account has a usage tier that governs rate limits (requests per minute, tokens per minute). New accounts start at Tier 1, which is enough for prototyping but tight for real production traffic. Consistent usage automatically moves you up tiers over months. You can also request a tier bump in Settings → Limits when you have a real deployment coming online.
The tier is not a spending limit — it's a rate limit. You can have $500 of credits at Tier 1 and still hit 429s if you burst too many requests in a minute. Tier 3+ is the sweet spot for most production workloads.
Workspaces and cost centers
Larger accounts organize keys into workspaces — one per project, team, or environment (dev / staging / prod). Each workspace has its own spending cap, its own keys, and its own cost breakdown in reporting. If your organization needs one API bill split across three products, workspaces are the tool. New accounts get a single default workspace and can add more anytime.
The four billing surprises that catch new accounts
- Output tokens cost 5x input tokens on every model. A chatty workflow that produces long responses costs disproportionately more than one that reads long inputs and outputs short ones. Watch response length before watching prompt length.
- Prompt caching is a 90% discount but only if enabled. Requests without `cache_control` pay full input-token rate every time, even on repeated identical prefixes. Turning caching on is one config change per request.
- Extended thinking bills as output tokens. If you enable adaptive thinking on Opus 5 at high effort, expect the model to spend a few thousand thinking tokens on a hard problem — invisibly, at output-token rates. Not free, not a separate line item.
- Failed requests still bill on retries. If your code retries on 5xx errors, each retry that returns any response bills for the tokens processed. Rate-limit-aware retry with backoff, not blind retry loops.
How to actually see where your money went
The Settings → Usage tab in the console shows daily spend broken down by model and by workspace. For deeper reporting (per-key breakdown, per-endpoint costs), the Admin API exposes the same data as JSON — automate a weekly export if you're spending enough that per-key attribution matters. Most solo developers don't need the Admin API; teams with more than a few workflows in production usually do.