OpenAI has introduced GPT-4o Mini, a more affordable and lightweight model for developers. This new model is significantly less expensive than full-size models and is touted as more capable than GPT-3.5. With this model, OpenAI aims to make artificial intelligence (AI) technology more accessible to a wider audience.
Developers can face high costs when developing applications using OpenAI’s models. This is a major hurdle, especially for lower-budget projects. Many developers have had to resort to cheaper models like Google’s Gemini 1.5 Flash or Anthropic’s Claude 3 Haiku. OpenAI is entering the market with GPT-4o Mini as a solution to this problem.
“GPT-4o Mini truly fulfills our mission to make AI technology more accessible to the masses. If we want AI to benefit every industry, every application, and every corner of the world, we need to make AI technology much more affordable,” said Olivier Godement, OpenAI API platform product leader.
Wide range of uses for GPT-4o Mini
Starting today, ChatGPT users will be able to use GPT-4o Mini on the Free, Plus, and Team plans. This new model will replace GPT-3.5 Turbo, with Enterprise users getting access next week. GPT-3.5 will no longer be an option for ChatGPT users, but will continue to be available to developers via the API. Godement noted that they have not yet determined when GPT-3.5 will be removed from the API.
On the other hand, the new lightweight model will support text and images in the API and will soon support all multimodal inputs and outputs, i.e. video and audio. These features can be used to develop more capable virtual assistants. For example, assistants that understand your travel plans and make suggestions could be possible with this model. However, it is important to note that the model is designed for simple tasks and is not used to develop complex applications like Siri.
GPT-4o Mini achieved an 82 percent score on the benchmark exam Measuring Massive Multitask Language Understanding (MMLU). This exam consists of approximately 16,000 multiple-choice questions covering 57 academic subjects. When it was launched in 2020, the models performed quite poorly on this exam. GPT-3.5 scored 70 percent on this exam, while GPT-4o scored 88.7 percent. Google claims that Gemini Ultra achieved the highest score ever on this exam, 90 percent. Comparatively, the rival models Claude 3 Haiku scored 75.2 percent and Gemini 1.5 Flash scored 78.9 percent.
It is important to note that the way benchmark tests are implemented varies slightly from company to company, making it difficult to compare scores from different models. Another important consideration is that the AI has these answers in its dataset, which can lead to cheating on the exam. Another important issue is that third-party evaluators are usually not involved in the process.
Developers are looking to reduce costs when building AI applications. The launch of GPT-4o Mini gives these developers a new tool. OpenAI gave fintech startup Ramp the opportunity to test the model. Ramp developed a tool that extracts spending data from receipts using GPT-4o Mini. Instead of filling in text boxes, users upload photos of their receipts, allowing the model to edit this data. Email client Superhuman also tested GPT-4o Mini and improved its ability to generate automatic reply suggestions.
This new model aims to give developers the opportunity to create applications and tools that they could not with larger and more expensive models. Many developers turned to the Claude 3 Haiku or Gemini 1.5 Flash models to avoid the high processing costs.
As a result, OpenAI’s launch of GPT-4o Mini is seen as a significant step forward in bringing AI technology to a wider audience. “I think this model will be very popular. It will be used by both existing AI-enabled applications and many applications that previously could not access this technology due to price,” Godement said.