GPT-3.5 (text-davinci-002/003)

What it is

GPT-3.5 is the label OpenAI used for the generation of large language models that sits between GPT-3 and GPT-4. It was not introduced in a single announcement: the name covers a series of API model versions released over roughly a year, beginning with text-davinci-002 and code-davinci-002, which OpenAI made available in its API on 15 March 2022. Those versions came with edit and insert capabilities and were trained on data up to June 2021^[1].

The series is best understood as a lineage. OpenAI describes GPT-3.5 as the successor of GPT-3 and Codex, and as the model family followed by GPT-4 in March 2023^[1]. The two first versions carry that history in their names: text-davinci-002 is a text model of the new generation, while code-davinci-002 is the same generation applied to code.

How the series was built out

The edit and insert capabilities were the visible change for developers. Instead of only continuing a piece of text from its end, the model could be handed an existing passage and asked to rewrite part of it or to add material in the middle.

A second text version, text-davinci-003, was introduced on 28 November 2022^[1]. ChatGPT followed on 30 November 2022. OpenAI described ChatGPT as fine-tuned from a model in the GPT-3.5 series that finished training in early 2022^[2], and as a sibling model to InstructGPT, which is trained to follow an instruction in a prompt and provide a detailed response^[5]. During its research preview, the blog said, ChatGPT was free to use^[5]. Both ChatGPT and GPT-3.5 were trained on Azure AI supercomputing infrastructure^[2].

For developers, the most consequential member of the family was gpt-3.5-turbo. It was chat-optimised and priced at one-tenth the cost of text-davinci-003, and OpenAI described it as the most capable and cost-effective GPT-3.5 model^[1].

Access to fresher information arrived later in the form of GPT-3.5 with Browsing (alpha), which was trained on data up to September 2021, introduced on 10 April 2023, and made available to ChatGPT Plus users on 27 April 2023^[1].

Where it sits in the line

The public face of the family was ChatGPT, but the API versions were what developers built on, and their lifetimes were set deliberately. In ChatGPT, GPT-3.5 remained the default free model until GPT-4o mini replaced it in July 2024^[4]. In the API, the older models were retired: text-davinci-002, text-davinci-003 and code-davinci-002 were shut down on 4 January 2024^[3].

Read as a whole, GPT-3.5 marks the point at which OpenAI's models moved from being asked to continue text to being asked to follow instructions and hold a conversation. Its two best-known forms — the davinci API models with edit and insert, and the chat-optimised turbo model at a tenth of the price — show how quickly the unit of use shifted from a single completion to a dialogue, and how much of the family's later role was defined by cost as well as capability^[1].