Anthropic API billing explained

Claude API usage is metered in tokens, and most organizations pay Anthropic for it ahead of time, with prepaid credits. This post walks through paying, token prices, the two big ways to pay less, spend caps and invoices. Every figure is as listed on the linked page when we checked in October 2026. Anthropic does change prices, so look again before you rely on one.

Prepaid credits or a monthly invoice

Anthropic's billing article describes two setups. Most organizations pay with prepaid usage credits. Organizations with an invoicing arrangement through Anthropic's Sales team are billed monthly in arrears instead.

Prepaid credits

Credits are bought on the Billing page of the Claude Console, by someone with an Admin or Billing role. Purchased credits are available right away, and they cover API access, playground usage and Claude Code.

The same article sets the rules. You're billed only for successful API calls, and failed requests aren't charged. There is one edge case: a request that was on track to succeed is still charged if your client disconnects or times out partway, which matters if your client timeouts are short. Let the balance hit zero and API calls stop until you add more. Credits expire one year from the purchase date, and you can't get a refund on a purchase.

Auto-reload

Auto-reload runs off two numbers you choose, a threshold and a target. When the balance falls below the threshold, it buys enough credits to bring you back up to the target. Turn it on while your balance is already at or below the threshold and you're charged right away.

The point of auto-reload is that production traffic doesn't stop on an empty balance. Because it keeps buying, pair it with a spend limit (more on those below) if you want usage to stop at a number you picked.

Monthly invoicing

On monthly invoicing, Anthropic adds up usage across API calls, Console usage and other services and bills it at standard pay-as-you-go pricing. The invoice comes from Stripe at the end of each calendar month.

Credits that come with Max and Team plans

Claude Max and Team plans include monthly Claude API credits once you link a Console organization. Per the billing article, these are spent before purchased credits, and unused ones expire at the end of each billing cycle. They can't be used in Claude Code or the Claude apps.

What a token costs

The pricing page lists five rates per model: base input tokens, 5-minute cache writes, 1-hour cache writes, cache hits and refreshes, and output tokens.

Claude Sonnet 5, for example, is listed at $2 per million input tokens and $10 per million output tokens. On most models a cache hit costs 0.1x the base input price, but there are exceptions. The pricing page notes that Claude Fable 5.1 and Claude Mythos 5.1 use 0.025x, and Claude Opus 5.5 and Claude Sonnet 5.5 use 0.05x. That works out to $0.20 per million cached tokens on Opus 5.5 and $0.10 on Sonnet 5.5.

One line on the pricing page is easy to overlook. Claude 4.7 and later models use a newer tokenizer, which the page says "produces approximately 30% more tokens for the same text." The same prompt can come out to more tokens on a newer model, so a cheaper per-token rate won't always mean a cheaper bill. Compare cost per task when you switch.

Two ways to pay less

Prompt caching

Prompt caching lets a request resume from a prompt prefix that's already cached. The simplest setup is one top-level cache_control field (Anthropic calls this automatic caching). For finer control, put cache_control on specific content blocks instead.

The cache covers tools, system and messages, in that order, up to the marked block. Anthropic's advice is to front-load whatever doesn't change: tool definitions, instructions, context, examples. A cached prefix lives for 5 minutes by default, and each use of the cache refreshes it at no additional cost. A 1-hour duration is available if you pay extra.

Caching helps with rate limits as well. Per the rate limits page, tokens read from cache don't count toward input tokens per minute limits for most models.

Message Batches API

The Message Batches API processes requests asynchronously, and the pricing is simple: "All usage is charged at 50% of the standard API prices." Most batches finish in less than 1 hour. A batch expires if processing doesn't finish within 24 hours.

A single batch holds up to 100,000 requests or 256 MB, whichever comes first. Results are available for 29 days after creation. Batches can also go slightly over your Workspace's spend limit, so don't treat that limit as an exact stop for batch work.

If the requests in a batch share context, Anthropic suggests the 1-hour cache duration, since batches can take longer than 5 minutes.

Spend caps and usage tiers

Each usage tier has a monthly spend cap, set on the rate limits page:

Usage tier	Monthly spend cap
Start	$500
Build	$1,000
Scale	$200,000

That is a big jump between Build and Scale. Custom tier organizations have no monthly cap. Anthropic places organizations on a tier automatically, based on usage history and account standing. Reach your tier's cap and API usage pauses until 00:00 UTC on the first day of the next month, unless you get a higher limit sooner.

You can also set a lower limit of your own under Settings > Billing in the Console. Once usage reaches it, requests return an error until access resumes or you raise or remove the limit.

Invoices and receipts

Anthropic's invoices article explains what you receive and where. Prepaid organizations get a receipt for each credit purchase, auto-reload purchases included. Organizations on a paid usage contract get a Stripe invoice at the end of every calendar month.

Each invoice is emailed to your billing email address. To find them, search your inbox for "Your receipt from Anthropic." Admin and Billing users can also download invoices from Invoice history on the Console Billing page.

Two smaller rules. Invoices below $0.50 aren't charged on their own; the amount rolls into the next invoice as an applied balance. And issued invoices can't be changed, so update billing details before the next charge.