OpenAI has added a new large language model (LLM) called GPT-4o mini to ChatGPT and its APIs. As the name implies, the GPT-4o Mini model is a smaller version of the GPT-4o model introduced in May. The mini model is designed to balance the power of GPT-4o with a more cost-effective approach.
GPT-4o mini has much of the functionality of its larger cousin, though the API only has text and vision support for now, with image, video, and audio inputs and outputs still in development. Like GPT-4o, the new model has a context window of 128,000 tokens, or eight times that of GPT-3.5 Turbo. The new model also comes with enhanced security features. Along with those already built into GPT-4o, GPT-4o mini has added new techniques that make it more resistant to jailbreaks and bad message injections, among other issues that plague developers looking to deploy AI APIs broadly.
Ready for bigger jobs
OpenAI suggests that the larger context window and other improvements, such as better understanding of non-English text, will make GPT-4o mini especially useful for processing large documents or linking multiple interactions with the AI model. For example, it could provide better recommendations in online stores, speed up real-time text responses for customer service, and produce accurate, detailed answers for students studying for a test more quickly than other models. OpenAI envisions GPT-4o automating and streamlining business processes thanks to its ability to source data and take action with external systems. For businesses using the API, the cost is significantly reduced to just over half the per-token price of GPT-3.5 Turbo.
“OpenAI is committed to making intelligence as accessible as possible,” OpenAI explained in its announcement. “We expect GPT-4o mini to significantly expand the range of applications built with AI by making intelligence much more affordable.”
GPT-4o mini is part of a recent wave of smaller LLMs, such as Google’s Gemini Flash and Anthropic’s Claude Haiku. However, according to OpenAI, GPT-4o mini outperforms them on many of the standard tests. The model scored 82% on the Massive Multitask Language Understanding (MMLU) benchmark, compared to 77.9% and 73.8% for Gemini Flash and Haiku, respectively. The same is true for the MGSM and Human Eval tests, where GPT-4o Mini scored 87% and 87.2%, while Gemini Flash scored 75.5% and 71.5%, and Haiku scored 71.7% and 75.9%. In other words, GPT-4o Mini wins on text understanding as well as math and coding tasks, as you can see in the graph below.
Mini Model Main Plans
According to OpenAI, the introduction of GPT-4o Mini represents an important step towards making advanced AI more affordable and accessible. Lower costs and better performance will likely help incorporate AI into everyday applications. The same goes for ChatGPT users, who can access the model starting this week. OpenAI also has plans to introduce fine-tuning capabilities for GPT-4o Mini within the API.
The bigger picture shows another step in the evolution of ChatGPT services. As OpenAI phases out GPT-3.5 for ChatGPT, the focus shifts to the next stage of providing more powerful models. OpenAI CEO Sam Altman has long hinted at how GPT-5 will “substantially improve” existing models. At the same time, OpenAI’s leaked scale for measuring AI power shows that there is still a long way to go until the still-mythical artificial general intelligence (AGI) that can perfectly mimic the workings of the human mind.