GPT -4o mini was announced yesterday and it can be called a surprise from OpenAI. The new version of the AI model debuted without any major announcements. What are the characteristics of the next version of the most popular chatbot?

GPT-4 Mini – unlike GPT-3.5, GPT 4o , GPT-4 and DALL E – is intended only for developers. It is to be cheaper than the paid versions, more accessible and more advanced than GPT-3.5, which is soon to be completely withdrawn.

GPT–4o mini suddenly appears as a cheaper and more accessible option for developers

GPT-4o mini

As of yesterday (June 18, 2024), ChatGPT users in Free, Plus, and Team programs can use GPT-4o Mini instead of GPT-3.5 Turbo, with Enterprise users getting access next week. GPT-3.5 will soon be gone for regular users, but will still be available to developers via the API if they don’t want to switch to GPT-4o Mini. This is a significant step for OpenAI in building smaller AI models, and a response to competitor products like Llama 3 8b, Claude Haiku, and Gemini 1.5 Flash.

I think the GPT-4o Mini really plays into OpenAI's mission of making AI more accessible to people. If we want AI to benefit every corner of the world, every industry, every application, we need to make AI much more affordable

Olivier Godement in a short interview for The Verge

GPT-4 Mini performs better than other simple AI models available on the market

GPT-4o mini
photo: OpenAI

GPT-4 Mini is expected to be up to 60% cheaper than GPT-3.5 Turbo and GPT 4o . OpenAI says the cost is around 15 cents per million input tokens and 60 cents per million output tokens. For now, the model is supposed to support text and vision, but developers also plan to introduce video, audio, and image files. The model's knowledge cutoff date is October 2023.

This is a huge advantage for developers who want to easily and cheaply create AI tools through the API. OpenAI claims that GPT-4 Mini outperforms other simple AI models on the market.

Devin is looking for a job on reddit. AI programmer posted a thread himself and listed his salary

These are not empty words, as the AI model was tested in the LMSYS.org chatbot arena. It scored 82% MMLU . This is a measure of massive multi-tasking language comprehension. It consists of about 16,000 multiple choice questions covering 57 academic subjects, including mathematics, philosophy, law, and medicine. For comparison – Claude 3 Haiku scored 75.2%, and Gemini 1.5 scored 78.9%.

Source: The Verge , OpenAI