OpenAI’s GPT-4o Mini AI Model Gives ChatGPT a Big Upgrade

OpenAI has launched GPT-4o mini, a lightweight version of its flagship AI model that should make the technology more accessible and give ChatGPT an important boost.

The new large language model costs about 60% less than GPT-3.5 Turbo for developers at 15 cents for a million input tokens and 60 cents per million output tokens.

However, it’s also billed as more capable than Turbo and some of the competition. GPT-4o mini reportedly outperforms GPT-3.5 Turbo, Google’s Gemini Flash, and Anthropic’s Claude Haiku in reasoning, math (and coding), and multimodal benchmarks. While these won’t always reflect real-world results, they suggest the improvements are across the board.

Mini currently only processes images and text. It will ultimately handle audio, video, and other content types, according to OpenAI. It has the same safety measures as the main 4o model, including efforts to avoid training on toxic material like hate speech and misinformation. GPT-4o mini is the company’s first model to use a new instruction approach that limits jailbreaks and other attempts to skirt around the rules.

GPT-4o mini is available now across OpenAI’s programming kits. More importantly, though, it just replaced GPT-3.5 Turbo in the free and paid ChatGPT clients.

The move won’t eliminate the mistakes that ChatGPT sometimes makes, and companies determined to get the most capable AI will still want to turn to the regular GPT-4o. It should raise the baseline quality of the prompt-based generator, though, and could improve what you get from Apple Intelligence and other apps or platforms that lean on ChatGPT for some of their functionality.

This also steps up competition between OpenAI and its main rivals Anthropic and Google. While much of the attention has been given to full-size AI models, OpenAI’s testing suggests lighter versions are increasingly good enough for many people. That’s important as AI moves from the internet to on-device models.