GPT-4 API is now generally available to paying customers

The GPT-4 API has become generally available to all paying customers as older models are set to retire in the coming months.

The ChatGPT API was introduced in March and millions of developers have requested access in the months since, according to OpenAI. As of 6 July 2023, all existing paying customers can access the GPT-4 API with 8K context while access to new developers will be opened by the end of the month.

Rate-limits will be raised after that, depending on compute availability.

The move from the initial Completions API - introduced in June 2020 - to the Chat Completions API has been a success for OpenAI. According to the company, the Chat Completions API now accounts for 97% of its API GPT usage.

Alongside the general availability of the GPT-4 API, older models using the Completions API are set to be deprecated and the API labelled as ‘legacy’, although OpenAI said it would still remain accessible.

Older completion models, including ada, babbage, curie and davinci will no longer be available as of 4 January 2024 and be automatically replaced by ada-002, babbage-002, curie-002 and davinci-002. Other older completion models, for example text-davinci-003, will require a manual upgrade to gpt-3.4-turbo-instruct.

The new models are due to become accessible in the coming weeks for testing ahead of the upgrade in 2024.

Older embedding models (for example text-search-davinci-doc-001) will also be retired by 4 January 2024. OpenAI released text-embedding-ada-002 in December 2022 and it has since accounted for 99.9% of all embedding API usage.

OpenAI said that it recognized that the move represented a “significant change for developers using those older models” and choosing to wind down these models “is not a decision we are making lightly”.

The company said it would be in touch with affected users, adding “we will cover the financial cost of users re-embedding content with these new models”.

Moving to Chat Completions

Initially a freeform text prompt was used for interaction with the company’s language models. However, OpenAI believes better results can be achieved using a more structured prompt interface.

The company claims the Chat Completions API structured interface and multi-turn conversation capabilities will make for a better conversational experience while also lowering the risk of prompt injection attacks.

Prompt injection attacks involve crafting an input prompt in a way to extract sensitive user information or exploit downstream systems that rely on the output of the language model.

Legal problems

OpenAI’s tools have proven popular with companies seeking to add generative AI to their products. However, the platform has also attracted the attention of regulators and lawyers.

The company has been the subject of multiple lawsuits, most recently being named in a complaint filed in the Northern District of California US District Court regarding the use of data in its models.

It is also being sued for defamation after the output of one of its tools erroneously accused a Florida radio host of financial claims.