OpenAI launches its API for accessing its models

Background

An application programming interface, or API, is a standard way for one program to request a service from another. For AI models, an API means developers can add a model's abilities to their own products by calling a service, without training a model themselves.

What happened

On 11 June 2020 OpenAI announced it was releasing an API for accessing new AI models that it had developed.^[1] On 22 September 2020 it agreed to license GPT-3 to Microsoft for Microsoft's own products and services.^[2]

Key details

By 25 March 2021 OpenAI said over 300 applications were delivering GPT-3-powered search, conversation, text completion and other features through the API.^[3]

On 10 August 2021 OpenAI released an improved version of Codex, its system that translates natural language to code, through the API in private beta.^[4]

On 18 November 2021 OpenAI made the API available with no waitlist, citing progress on safety.^[5]

On 27 January 2022 OpenAI made its InstructGPT models the default language models on the API. The InstructGPT training data included prompts submitted through the API.^[6]

At its first DevDay on 6 November 2023, OpenAI said more than two million developers were building tools with its API.^[7]

An API lets a developer send an input, such as a text prompt, to a service and receive the output in reply. The developer does not need to download the model or run it on their own hardware, and the provider can update or restrict the service from one place. This is different from publishing a model for anyone to run.

Access was not open to everyone at first. A waitlist is a queue in which people request access and wait for approval. The change in November 2021 ended that queue for the API.^[5]

Codex, mentioned above, is a system that turns a description written in everyday language into code. Its improved version was offered in private beta, a limited test in which only a selected group could try the service while it was still being developed.^[4]

Impact

Opening the API let developers build products on top of OpenAI's models without training their own, which separated the work of building models from the work of building applications. The InstructGPT detail shows that prompts sent through a hosted API can feed into later training, which is one reason hosted services raise questions about how submitted data is used.

For a developer, the practical effect was that a company could add language features to an application by writing a request, rather than collecting data and training a model from scratch. For the provider, a hosted service keeps the model under its own control, which is part of why the work of building models and the work of building applications came to be separated. Readers evaluating such services are best served by checking how each provider describes its handling of submitted data.