GPT-4o mini

What it is

GPT-4o mini is a small, low-cost language model released by OpenAI on 18 July 2024^[1]. It carries the GPT-4o name and shares the improved tokenizer introduced with GPT-4o, which OpenAI said made non-English text handling more efficient^[7]. The mini label marks it as the cheaper, smaller member of the family: in general a smaller model gives up some capability in exchange for a much lower price per token, which is why labs bring these models out alongside their flagships and aim them at high-volume work.

Price

Price was the headline. On the API it cost $0.15 per million input tokens and $0.60 per million output tokens, against $2.50 and $10 for GPT-4o; fine-tuning doubled those prices^[1]. The Decoder reported that the model scored 82% on the MMLU benchmark and was more than 60% cheaper than GPT-3.5 Turbo^[2]. MMLU is a multiple-choice test that samples questions across many academic subjects, and it is one of the standard numbers used to compare general knowledge across models. A cheap model that posts a solid score on it changes what developers can afford to build: labelling, extraction, routing and summarising need many calls rather than one very good one, and at this price level those calls become practical.

Capabilities

The model had a context window of 128K tokens^[5]. It supported up to 16,000 output tokens per request and had knowledge up to October 2023^[7]. At launch it handled text and vision through the API, and OpenAI said text, image, video and audio inputs and outputs were planned for the future^[7]. The context window matters in practice because it sets how much material — a long document, a transcript, a batch of records — can be handed to the model in a single request before the developer has to split it up; the output limit similarly caps how long one generated answer can be. A knowledge cutoff means the model does not know about anything after that date unless it is pasted into the prompt.

Where it sat in the lineup

GPT-4o mini replaced GPT-3.5 Turbo in ChatGPT on the day it launched^[1], which put the cheaper model in front of the product's users and not only in the API. Its time in that slot was short: in May 2025 GPT-4.1 mini replaced GPT-4o mini for all ChatGPT users^[6]. OpenAI's GPT-5 system card names gpt-5-main-mini as its successor^[4], so the mini line continued, but this particular generation was superseded twice.

Life on the API

For developers who pin a specific version, OpenAI published a dated snapshot, gpt-4o-mini-2024-07-18, and scheduled it to shut down on 14 January 2027^[3]. Versioned snapshots with a published retirement date let a team keep a stable target while the named alias behind it moves on to a newer model.

Why it matters

Two things make the model worth remembering. The first is the price gap: charging roughly an order of magnitude less than the flagship for a model that still scores respectably on a widely used benchmark is what made the small-model tier a serious place to build products, rather than a fallback for demos. The second is the turnover. A model introduced as the default in the chat product was replaced there within less than a year, and named as superseded again in a later system card^[6][4]. That is a useful picture of how fast the cheap end of the market moved: the interesting question for a developer is less which small model is best at any moment than how quickly the one they built on will be retired, and how much notice they get. The published shutdown date for the pinned snapshot is the part of the story that answers it^[3].