OpenAI opens the GPT-3 API in private beta

OpenAI opened a private beta of the OpenAI API, giving outside developers access to GPT-3 for the first time through a commercial, metered web service rather than as downloadable weights or open code. Access required applying for a waitlist; OpenAI reviewed applications and reserved the right to cut off any use case it judged harmful or abusive.

The company described the model as general-purpose — “text in, text out” — capable of tasks ranging from answering questions about a passage to generating creative writing, translating between languages, and writing simple code, all from the same underlying model steered only by the prompt. Early access went to a small set of partners; the game AI Dungeon was among the first products built on it, and OpenAI pointed to it and similar prototypes as evidence of the API’s breadth.

OpenAI framed the API delivery model explicitly as a safety choice. Rather than publishing the model’s weights, as it had done with smaller predecessors, the company argued that an API let it “broaden access over time” while retaining the ability to restrict or revoke use if problems emerged — a reversal of the publish-everything norm that had governed most prior language model research, including OpenAI’s own earlier, partial release of GPT-2.

The decision set a template that later defined much of the industry: GPT-3 itself was never released as open weights, and the API-as-product model — metered, reviewed, revocable — became the default distribution method for the largest frontier models from OpenAI, and subsequently Anthropic and Google, even as competitors including Meta and various open-source efforts pursued the opposite path. The waitlist itself lasted roughly a year and a half before OpenAI made the API generally available.