Access, not a download
OpenAI's announcement described the API as a general-purpose "text in, text out" interface, letting a developer try it on essentially any English-language task rather than calling a system built for one job. [1]
The company had published GPT-2's weights the year before. GPT-3 went the other way: nobody outside OpenAI could download the model. A developer sent text to an endpoint and got text back, with OpenAI holding the weights, the compute and the ability to cut off a use case it did not like. [2]
The waitlist doubled as a review process
Getting in required an application stating an intended use. Personal experimentation got fast access to the playground; a stated use like content generation drew closer review before OpenAI would grant it. [2]
Usage was metered per token. The most capable Davinci model cost about six cents per thousand tokens, and new accounts got roughly eighteen dollars of free credit to start with, spent over three months. [2]
Gatekeeping and billing arrived in the same product. A single signup form let OpenAI screen for the uses it wanted to allow and start a meter running on the ones it approved, which a weights release could not do. [2][1]
Microsoft got the weights, everyone else kept the endpoint
Three months later, Microsoft announced an exclusive license to GPT-3's underlying model for its own products. OpenAI said it would keep running its own Azure-hosted API for outside developers, unaffected by that deal. [3]
That split, one company holding the weights and everyone else buying calls to an endpoint, became the industry's default shape for the years that followed. The API launch is what made the split possible: there was already a product to carve exclusivity out of. [3]
The gate came down, but not for a while
The waitlist lasted about a year and a half. OpenAI dropped it in November 2021, letting developers in supported countries sign up directly instead of waiting for an approval email. [4]