OpenAI releases text-davinci-003 in the GPT-3.5 series

Background

In March 2022 OpenAI began the line of models it called the GPT-3.5 series with two entries, text-davinci-002 and code-davinci-002^[1]. Models of this kind are usually reached through an API: a developer sends text to the model, receives text back, and builds the surrounding product. In that setup the model is a component rather than a finished thing — how it is presented to a user is decided by whoever calls it.

What happened

On 28 November 2022 OpenAI introduced text-davinci-003, a new model in what it called the GPT-3.5 series^[1]. Two days later, on 30 November 2022, the company launched ChatGPT, fine-tuned from a model in the same series^[2].

Fine-tuning means taking a model that has already been trained on a large body of text and training it further so that it behaves in a particular way. For ChatGPT, OpenAI described the result as a model trained to follow an instruction in a prompt and provide a detailed response^[4].

Key details

OpenAI described ChatGPT as a sibling model to InstructGPT^[4]. During the research preview, ChatGPT was free to use^[4]. A research preview in this sense is a release made while a system is still being studied, so that its makers can see how it is used and where it fails.

text-davinci-003 stayed available as an API model. For most uses it was later superseded by gpt-3.5-turbo, which was priced at one-tenth of its cost^[1]. The older model was shut down in the API on 4 January 2024^[3].

What the dates show

Read as a sequence, the releases describe how one model line moved. The series opened in March 2022 with two models^[1]. text-davinci-003 joined it on 28 November 2022^[1], and ChatGPT appeared two days later — not a separate series, but a model of the same series fine-tuned for a new use^[2].

The difference most visible from outside is the shape of the interface. On one side a developer calls an API and gets text back; on the other, a person types an instruction and reads a detailed answer, free of charge during the research preview^[4]. The same underlying family of models therefore appeared in two quite different forms of access: one built for software to call, one built for a person to talk to.

Price and retirement are the other half of the story. gpt-3.5-turbo took over most of text-davinci-003's work at one-tenth the cost^[1], and text-davinci-003 was retired from the API on 4 January 2024^[3]. On this line, a newer model replacing an older one at a lower price, and the older one being switched off, can be observed directly rather than inferred. The two November releases also stand as a record of how short the gap can be between a model appearing in a developer-facing series and the same kind of model appearing as a free, instruction-following chat service^[2][4].