a response to Google and Anthropic’s lightweight models

As generative AI becomes mainstream, OpenAI launches GPT-4o Mini to make this technology more accessible.

Artificial intelligence giant OpenAI has unveiled GPT-4o Mini, a language model that is lighter and significantly cheaper than its big brothers. And this comes just weeks after GPT-4o and in the midst of GPT-5 rumors.

A little one who has everything of a big one

First surprise: GPT-4o Mini would be more efficient than GPT-3.5, the model that made ChatGPT’s heyday in its early days. We’re talking here a score of 82% on the famous MMLU (Measuring Massive Multitask Language Understanding) test, compared to 70% for GPT-3.5. For the uninitiated, imagine an ultra-tough exam with 16,000 multiple-choice questions on 57 different subjects.

But let’s not get ahead of ourselves. These benchmarks are a bit like comparing grades from different schools: not always very reliable. Some researchers are sounding the alarm, pointing out the differences in methods between companies and the risk of “cheating” if the AI ​​has already seen the answers in its training data. In short, let’s take these numbers with a grain of salt.

What is certain is that OpenAI does not do things by halves. GPT-4o Mini arrives with the whole package: support for text, vision, and soon video and audio.

But what is this LLM for?

The crux of the matter for many developers is the cost of AI models. With GPT-4o Mini, OpenAI clearly wants to appeal to those who cannot afford GPT-4’s services. The idea is to access a cheaper, but still effective, LLM.

OpenAI has already let a few companies test the toy. Ramp, a fintech startup, has created a tool that extracts data from receipts in a flash. No more hours of typing numbers, GPT-4o Mini does it for you. Superhuman, an email client, uses it to automatically suggest replies to emails. Handy, right?

By the way, GPT-3.5 is almost over. At least on ChatGPT. OpenAI has decided to replace this good old model with GPT-4o Mini for all users, whether they have a free or paid account. Don’t panic too quickly, developers addicted to GPT-3.5 will still be able to use it via the API for a while. But OpenAI warns: its days are numbered.

One might wonder why it took OpenAI so long to release a lightweight model, when Google and Anthropic already had their own protégés on the market. According to Brad Lightcap, OpenAI’s COO, quoted by The Vergeit was a matter of priorities. The company was too busy building ever bigger and better models, like GPT-4.