OpenAI releases GPT-4o mini

Background

A language model is billed by the token — the fragments of text, roughly parts of words, that the model reads as input and writes as output. Providers charge per token because every token costs computing power, and a larger model costs more computing power per token than a smaller one. That is why a model family usually comes in several sizes, and why the cheapest model in a family often decides whether a service can afford to run at all.

What happened

On 18 July 2024, OpenAI released GPT-4o mini, a smaller and cheaper version of GPT-4o^[1]. It was not only a product for developers using the API: the same day, it replaced GPT-3.5 Turbo as the model behind ChatGPT^[1].

Key details

On the API, GPT-4o mini cost $0.15 per million input tokens and $0.60 per million output tokens^[1]. Reporting at the time described it as more than 60% cheaper than GPT-3.5 Turbo while scoring 82% on MMLU^[2]. MMLU is a benchmark of multiple-choice questions across a wide range of academic subjects, so the score is a rough measure of how much general knowledge a model can recall and apply, not of how well it holds a conversation.

OpenAI gave the same 82% figure and said the model outperformed GPT-4 on chat preferences on the LMSYS leaderboard at the time^[3]. That leaderboard is built from users comparing the answers of two anonymous models side by side, so it measures preference rather than correctness on a fixed test set. The two numbers therefore say different things: one about knowledge, one about which answer people tend to pick.

GPT-4o mini shipped with a 128K-token context window, and at launch the API accepted both text and images as input^[3]. The context window is the amount of material that can be put in front of the model in a single request — a long document, an excerpt of code, a transcript — and it is one of the features that separates a deliberately small model from a stripped-down one. Vision input works the same way: an image is passed to the model alongside text, and the model answers about both together.

Why it matters

The release did two things at once. The price set a public reference point for what a model with a 128K-token context window and image input costs at the low end^[1][3]. And the swap inside ChatGPT moved the model that ordinary users talk to from GPT-3.5 Turbo to GPT-4o mini, so GPT-3.5 Turbo no longer filled that role^[1]. Both changes point in the same direction: capability that had been sold as a premium tier was pushed down to the cheapest tier and into the default product.

For developers, the low price is what makes certain uses worth considering at all — running a model over a large batch of documents, calling it once per user, or embedding it inside a small tool. The trade-offs are equally part of the picture. The 82% MMLU score and the leaderboard preference result were the figures given at release, and each describes only one aspect of a model; neither replaces trying it on the task you actually have.