# What an LLM API Is — Claude API / OpenAI API Basics

Source: https://www.geekswithgeeks.com/en/llm-apis/b-what

> See the API as a stateless web service that turns messages into a reply.

## A web service with one main call

A hosted language model is reached through a **web API**: you send an HTTPS `POST` with a JSON body (model name, **messages**, settings) and an **API key**, and receive JSON containing the model's reply and **usage** (how many input and output tokens were used). The service is **stateless**: it remembers nothing between calls, so your application sends the whole conversation each time. Billing is per **token**, so every request has a price and a latency. The two most widely used families are **Anthropic's Claude API** (the *Messages API*) and **OpenAI's API** (*Chat Completions* and the newer *Responses* API). Their details differ, but the ideas in this course apply to both and to most other providers.

## Request in, message out

An LLM API is an HTTPS call: you send messages and settings, and get a message and usage numbers back.

![Four parts: key, request, response, usage.](assets/figures/llm-apis/section-1-map.svg) — Figure 1.1 — Key, request, response and usage.

## A very forgetful helpdesk

Each call is a fresh phone call to a helpdesk agent who remembers nothing. You must repeat the relevant history every time you call.

## The shape of a call

Both providers follow this pattern, with different field names.

```text
POST  https://<provider-host>/<messages-endpoint>
headers:  API key (+ content-type, + version header for some providers)
body:     { "model": "<model-name>",
            "messages": [ {"role": "user", "content": "..."} ],
            "max_tokens": ..., "temperature": ..., ... }

response: { "content"/"choices": [ the reply ],
            "stop_reason"/"finish_reason": "...",
            "usage": { input_tokens, output_tokens } }
```

**Quiz:** What does "stateless" mean for an LLM API?

- [ ] It only works offline
- [ ] It stores your conversation forever for free
- [x] It remembers nothing between calls, so you resend the conversation
- [ ] It has no billing

*Answer:* It remembers nothing between calls, so you resend the conversation. Each request must carry all the context the model needs.