Updated October 15, 2026. This article preserves the context of GPT-4o mini’s July 2024 launch, with later product changes identified separately. It is a documentation-based review, not a hands-on performance test. The cost example below is fictional.

Featured illustration: retained historical conceptual artwork showing a stylized robot and circuitry; it is not a product screenshot or performance evidence.

API lifecycle update: OpenAI announced the deprecation of gpt-4o-mini-2024-07-18 on October 14, 2026, with API shutdown scheduled for January 14, 2027. Its fine-tuned versions follow the same schedule. These dates come from the official deprecation notice; a deprecated model is not the same as an already shut-down model.

What changed in July 2024

OpenAI introduced GPT-4o mini on July 18, 2024 as a lower-cost small model for text and image-input tasks. Launch pricing was US$0.15 per million input tokens and US$0.60 per million output tokens. OpenAI reported an 82% MMLU score and described the model as more than 60% cheaper than GPT-3.5 Turbo. Those were launch-era comparisons, not evidence that it remains the cheapest or strongest model available today. See OpenAI’s launch announcement.

The practical question was whether inexpensive calls could handle a narrowly defined job well enough to be useful. For a team sorting messages or drafting short answers, the relevant result would be how often the output needed correction. A benchmark score alone could not answer that question.

For broader historical context, our GPT-4o overview covers the larger model’s launch and subsequent changes. It is background for understanding the family, rather than a recommendation to choose either model for a new project.

Capabilities and token limits

Checked on October 15, 2026, the GPT-4o mini model page lists the alias gpt-4o-mini and dated snapshot gpt-4o-mini-2024-07-18, with these specifications:

  • Text and image inputs, with text output
  • A 128,000-token context window and up to 16,384 output tokens
  • A knowledge cutoff of October 1, 2023
  • Support for function calling and Structured Outputs
  • No audio or video support for this base model

The same page lists text-token rates of US$0.15 per million input tokens, US$0.075 for cached input, and US$0.60 for output. A listed capability does not override a shutdown notice or establish access for a particular API account.

A 128,000-token context window should not be converted to a fixed number of book pages. Tokens are not pages, and their relationship to words varies with language and encoding. Count the actual material, including instructions and conversation history, and allow room for the answer. OpenAI’s token-counting guide explains the distinction. A large context window also should not be read as permanent memory or a promise that every detail will be retrieved correctly.

A small workflow and cost example

Imagine a fictional shop evaluating a text-only assistant that labels support messages and drafts replies for staff. It processes 10,000 messages, each using 800 input tokens in total and producing 150 output tokens. Assume one call per message, uncached input, and the listed input/output rates above.

  • Input: 8 million tokens × US$0.15 = US$1.20
  • Output: 1.5 million tokens × US$0.60 = US$0.90
  • Illustrative model-token total: US$2.10

This is arithmetic, not a measured workload or an operating-cost quote. It excludes retries, extra calls, image inputs, tools, hosting, taxes and staff review. Longer prompts or answers change the total. The example demonstrates how to estimate one bounded workload; it does not establish that GPT-4o mini is the best option for it.

A useful pilot would separate straightforward delivery questions from refund disputes. Staff could compare draft replies with the shop’s actual policy, record incorrect promises, and measure editing time. If cheap drafts create substantial cleanup, the low token bill is only a small part of the decision.

Limits worth testing

Image input does not imply dependable document extraction. OpenAI’s vision guidance warns about small or rotated text, difficult chart elements, inaccurate descriptions and approximate object counts. A receipt workflow should test blurred photos and unusual layouts, then verify amounts against the original image.

OpenAI’s safety guidance also recommends testing adversarial inputs and having people review outputs, particularly in consequential applications. For the fictional shop, that means checking whether a customer’s message can redirect the assistant and keeping approval with staff before a reply makes a commitment. Built-in safeguards are not a substitute for those checks.

What changed after launch

ChatGPT: The July 2024 rollout is historical. OpenAI’s model release notes record GPT-4.1 mini replacing GPT-4o mini in ChatGPT on May 14, 2025. That replacement is not a statement about which model a present-day account uses, nor does a ChatGPT change determine the API shutdown date.

Fine-tuning: By its August 20, 2024 announcement, OpenAI documented GPT-4o mini fine-tuning for paid API tiers. The later self-serve fine-tuning timeline barred organizations without prior fine-tuning use from creating jobs on May 7, 2026. From July 2, job creation required fine-tuned-model inference in the preceding 60 days. New jobs for active existing customers are scheduled to end on January 6, 2027.

Audio and video: The launch announcement discussed future multimodal expansion. That roadmap should not be converted into a promise that the base model supports those formats. Check the exact model identifier and its documented modalities.

The lasting lesson

GPT-4o mini’s launch offers a useful case study in evaluating lower-cost AI: define the job, estimate its complete workload, test failures and include human review. For existing integrations, the announced retirement makes migration planning part of that evaluation. For new work, compare currently supported options using your own representative tasks.