API call
An API call is a request that one program sends over the internet to another program in order to retrieve a service from it. For AI services, the API call is the standard way to use a model — and the unit by which usage is billed.
Programs often need to do things they can’t do themselves. A weather app, for instance, doesn’t have its own weather stations. So it sends a request over the internet to a foreign computer that has the weather data. This request is called an API call. API stands for “Application Programming Interface,” meaning a fixed interface through which programs talk to each other. The call consists of a clearly prescribed question, and back comes an equally clearly prescribed answer.
Why AI companies make money from it
Large language models don’t run on your laptop. They need expensive specialized chips in large data centers. Providers like OpenAI, Google, or Anthropic therefore make their models available via API. Every call travels to the data center, gets computed there, and comes back as a response.
The API call is thus also the cash register. Billing is usually based on tokens, i.e., text chunks roughly one syllable long. You pay for the text sent in and for the text generated, often in fractions of a cent per thousand tokens. Individually, that’s tiny. With millions of calls per day, this adds up to the revenues reported in business news.
For companies, this has a practical advantage. They don’t have to train their own model or buy hardware. They essentially rent the capability per use. The downside: anyone who builds their product on this is dependent on the price and availability of an outside provider.
What actually happens during a request
An API call follows a fixed pattern. The requesting program sends a message to an internet address, the so-called endpoint. The message contains the actual data — for AI services, that’s the text of the question and settings like the desired model. Added to this is a key, a kind of digital ID that lets the provider recognize whose account is being charged.
The data is transmitted in a uniform format, usually JSON. This is simply a structured notation that humans can read and machines can reliably parse. The response comes back in the same format and contains, besides the result, information about billing. Because everything is strictly defined, the connection works between programs that know nothing about each other.
Timing matters here. A call to a language model often takes one to several seconds, because the model computes the answer piece by piece. Providers also limit how many calls are allowed per minute. This limit is called a rate limit and prevents individual users from overloading the servers.
Where these interfaces show up in everyday life
Almost every app that displays current information works with API calls. Timetables, map services, exchange rates, payment processes in online shops: in all these cases, one program is querying another. You don’t notice any of this because the process runs in the background.
It’s the same with AI products. When a writing app summarizes texts or a coding tool suggests code, there’s almost always a call to a foreign model behind it. A common misconception is that the AI resides within the app itself. Usually, the app is just an interface that passes requests along.
In the news, the term mainly comes up in connection with outages and price changes. If a major API goes down, hundreds of dependent services stop working at the same time. If a provider lowers its prices per million tokens, this is seen as a competitive move. Anyone who understands what an API call is can properly interpret such reports.
