API cost means you pay per request, measured in tokens, with no monthly fee

What "API cost" actually means

An API is the door one piece of software uses to ask another piece of software to do something. API cost is the fee for each trip through that door. When an app asks Claude to draft an email, Anthropic charges for that one request. When it asks ChatGPT to summarize a form submission, OpenAI charges for that one.

The phrase shows up in three places a business owner will see it: on the pricing page of Anthropic or OpenAI, on a software vendor's page as an "AI usage" or "credits" line, and on your own invoice if you have connected your own key to a tool.

Here is what it is not. It is not a monthly fee for having access. There is no base charge on the Claude or OpenAI API. If you make zero requests in a month, you pay zero. Both providers do ask you to prepay credit before you start, so you will see a balance go down as you use it. It is also not covered by a ChatGPT Plus or Claude Pro subscription. Those are separate products, which we covered in why your Plus subscription does not pay for the ChatGPT API.

Outside AI, "API cost" means the same thing for other services. Google Maps, address lookups and payment processors also charge per call. The rest of this page is about AI, because that is where the bill is most confusing.

How the bill is worked out, with real numbers

AI APIs charge by the token, which is a chunk of text. OpenAI's rule of thumb is that one token is about four characters of English, or roughly three quarters of a word. So 1,000 tokens is about 750 words.

Every request has two parts, and they are priced separately:

• Input tokens: everything sent to the model. That includes your instructions, the email you want answered and any business details the tool attaches.
• Output tokens: everything the model writes back.

Output costs more, usually four to eight times more per token. The table uses list prices from the providers' pricing pages when we checked. They change, so check the current page before you budget.

*One reply means 1,500 input tokens (the customer's email thread, your instructions and some business details) and 300 output tokens (a reply of about 225 words). For Claude Sonnet the math is 1,500 × $3 ÷ 1,000,000 = $0.0045 for input, plus 300 × $15 ÷ 1,000,000 = $0.0045 for output. That totals less than a cent.

That is the answer most people are looking for. For ordinary small-business writing, one AI request costs a fraction of a cent to about a cent. For full monthly estimates by type of business, see what Claude actually costs a small business.

Why the same task can cost ten times more than you expect

The per-request number stays small only if the request stays small. Three things push it up, and none of them are obvious from the pricing page.

1. Chat history is re-sent every turn. The API does not remember the conversation. Each new message sends the whole conversation so far as input again. Ask ten follow-up questions in one chat and the tenth request carries all nine before it. A long back-and-forth can use far more input tokens than ten fresh requests would. Starting a new conversation for a new task is the cheapest habit you can have.

2. Attachments are input. A 20-page PDF contract is roughly 10,000 words, or around 13,000 tokens. On Claude Sonnet that is about four cents each time you send it, before the model writes a word. If you ask three questions about it in one chat, you pay for those tokens three times.

3. Model choice is a 5x to 10x lever. The table shows it. Claude Sonnet costs about nine times more per reply than GPT-5 mini for the same job. A short reply to a booking inquiry rarely needs the bigger model. Proposals, contract reviews and anything a client will judge you on often do.

To see your own number, open the usage page in your Anthropic or OpenAI console after a week of normal work. It shows tokens in and tokens out per day. Divide the dollar total by the number of tasks you ran and you have your real cost per task, with no guessing. Set a monthly spending limit on the same page while you are there.

How to read "AI credits" on a software pricing page

This is where knowing what API cost means pays off. Many business tools do not show you token prices. They sell "AI credits" or "AI actions" in bundles, and the tool pays the API cost behind the scenes. Nothing is wrong with that, but it hides the markup.

The test is one division. Here is a hypothetical example, not a real vendor:

• A tool sells 500 AI credits for $20, and one credit drafts one email reply. That is $0.04 per reply.
• The same reply on Claude Sonnet at list price is about $0.009.
• So you are paying roughly 4.4 times the raw API cost.

Run the same math on any tool you are considering. Find out what one credit buys, estimate the tokens with the 750-words-per-1,000-tokens rule, and compare it with the table above. Some vendors will come out close to list price because they are adding real work around the model. Others will be 5x or 10x. Either way you are now comparing prices instead of guessing.

The alternative is a tool that lets you plug in your own API key, so the API cost lands on your own Anthropic or OpenAI bill at list price. Fynl works this way, and so do a number of other tools. If you only use AI a few times a week, a credit bundle you never run out of may be simpler and the markup may not matter. If you are drafting dozens of replies a day, your own key is usually the cheaper route.

Questions people also ask

Sources

• Anthropic API pricing
• OpenAI API pricing
• OpenAI Help Center: What are tokens and how to count them