The OpenAI API: Access to Foundational AI Models

WHY IT EXISTS Training a frontier generative model costs many millions of dollars and requires infrastructure most companies will never build themselves. The OpenAI API exists to separate those two problems: OpenAI trains and hosts the model, and any developer can rent access to its capabilities by the request, without owning a GPU cluster or a research team.

THE MENTAL MODEL Treat it like electricity metering rather than a software license. You do not install GPT-4 or DALL-E, you send a request over HTTPS and pay for what you consumed, measured in tokens for text models and per image for DALL-E. The model always runs on OpenAI's infrastructure, your code just calls it, the same way a web app calls a payments processor instead of implementing card networks itself.

HOW IT WORKS A client sends a JSON request to an endpoint, most commonly the chat completions endpoint, authenticated with a bearer API key in the request header, specifying a model name, a list of messages, and parameters such as temperature and max tokens. The API returns generated text as JSON, or streams it back token by token for a responsive UI. Separate endpoints cover image generation through DALL-E, speech to text through Whisper, and vector embeddings for search and retrieval. Usage is billed per token consumed, input and output priced separately, and every account is subject to rate limits measured in requests and tokens per minute that scale with usage tier.

WHEN IT MATTERS The API is the right choice whenever a product needs generative capability without a research budget: drafting text, summarizing documents, generating images, powering a support chatbot. It matters less, or becomes the wrong default, when data cannot leave your own infrastructure for compliance reasons, or when cost at scale makes a smaller self-hosted open model cheaper for one narrow task.

ONE CONCRETE EXAMPLE A note-taking app adds an improve writing button. On click, it sends the user's paragraph to the chat completions endpoint with a system prompt asking for a clearer rewrite, streams the response back into the text box as it arrives, and the request costs a fraction of a cent based on the token count of the paragraph.