GPT-3.5 Turbo

What GPT-3.5 Turbo is

GPT-3.5 Turbo is a member of OpenAI's GPT-3.5 family of language models, tuned for conversation. OpenAI released it through its API on 1 March 2023, and it was the model behind what the company then called the "ChatGPT API"^[1][6]. It was not a separate product line so much as the same kind of model that powered ChatGPT, offered to developers as a paid endpoint their own software could call.

The price was the headline

At launch the most quoted number was the cost: $0.002 per 1,000 tokens, which OpenAI said was 10 times cheaper than its existing GPT-3.5 models^[1]. It has been described as the most capable and cost-effective GPT-3.5 model, optimised for chat, at one-tenth the cost of text-davinci-003^[2].

OpenAI attributed the reduction to system-wide optimisations that had cut the cost of running ChatGPT by 90 percent since its December 2022 release, savings the company said it passed on to API users^[7]. Unit price matters more than it might seem: a model that is cheap per token is one developers can afford to call thousands of times inside a single feature, which is what a conversational product usually requires.

Instacart, Snap and Quizlet were among the first customers of the ChatGPT API^[6].

A different shape of input

Unlike older GPT models, gpt-3.5-turbo processed a sequence of messages with metadata instead of unstructured text^[7]. In other words, a caller did not hand the model one block of prose and ask for a completion; it sent a structured list of messages, which is a closer match to how a chat assistant actually works.

The Decoder reported that early testers migrated from text-davinci-003 with only minor prompt changes^[7]. That mattered for adoption: teams already running on the older model could switch without rebuilding their applications from scratch.

Versions and retirement

At DevDay on 6 November 2023, OpenAI released an updated GPT-3.5 Turbo with a 16K context window by default, meaning each call could carry considerably more text before the model ran out of room^[3]. Its last stable snapshot, gpt-3.5-turbo-0125, was released on 25 January 2024^[2].

Models in the API are offered as dated snapshots, and OpenAI publishes a schedule for shutting them down so that teams have time to move to newer versions. gpt-3.5-turbo-0613 and gpt-3.5-turbo-16k-0613 were shut down on 13 September 2024. gpt-3.5-turbo-1106 and gpt-3.5-turbo-instruct were scheduled for 28 September 2026, and gpt-3.5-turbo-0125 for 23 October 2026^[5].

Where it sits in the line

GPT-3.5 Turbo was the chat successor of the GPT-3.5 series, and it was itself succeeded by GPT-4o mini as OpenAI's low-cost model^[4]. GPT-4o mini replaced GPT-3.5 Turbo in ChatGPT on 18 July 2024^[4]. Between those two points it served as the cheap, widely used default for developers who wanted a chat-capable model through an API.