OpenAI introduces GPT-4o mini: A more cost-effective and powerful AI model that will replace GPT-3.5. The multimodal system promises improved efficiency for developers and companies. OpenAI is expanding its portfolio of AI models and is presenting GPT-4o mini, a more compact and significantly cheaper version of its flagship model GPT-4o. The new language model is not only more cost-effective, but also more powerful than its predecessor GPT-3.5 Turbo, which it will replace in ChatGPT.
GPT-4o mini is particularly distinguished by its price-performance ratio. With costs of 15 cents per million input tokens and 60 cents per million output tokens, it is around 60 percent cheaper than the previous GPT-3.5 Turbo model. By comparison: the large GPT-4o costs more than thirty times as much at $5 per million input tokens and $15 per million output tokens. Despite the lower price, OpenAI promises improved performance. In the MMLU (Measuring Massive Multitask Language Understanding) benchmark for general knowledge and language comprehension, GPT-4o mini achieved a value of 82 percent, while GPT-3.5 achieved 70 percent. This positions the new model at the top of the smaller AI models, even if the gap to competing products such as Claude 3 Haiku from Anthropic or Gemini Flash from Google is not that great.
As a multimodal model, GPT-4o mini can process both text and images. OpenAI plans to integrate additional modalities such as video and audio in the future. The model’s context window includes 128,000 tokens, which is an eightfold increase over GPT-3.5 Turbo. Also notable is the ability to generate up to 16,000 output tokens – significantly more than most comparable models.
Another advantage of GPT-4o mini is its speed. At 166 tokens per second, it is one of the fastest language models available. This makes it particularly attractive for applications that require high processing speed.
OpenAI states that some companies are already using the model successfully. The fintech startup Ramp, for example, uses GPT-4o mini to extract output data from uploaded receipts. The email client Superhuman uses the model to suggest replies to emails.
With GPT-4o mini, OpenAI is also introducing a new security technique: the so-called “Instruction Hierarchy”. This method is intended to make it more difficult for users to circumvent predefined instructions or restrictions of the model by prioritizing certain instructions.
OpenAI says extensive security measures were implemented when developing the model, including filtering problematic content during training and applying techniques such as Reinforcement Learning with Human Feedback (RLHF). The new model is available now for users of ChatGPT’s Free, Plus, and Team plans. Companies with Enterprise plans will get access in the coming week. Developers who want to continue using GPT-3.5 can do so through the API for now, but OpenAI plans to remove GPT-3.5 from the API entirely at a later date.
With the introduction of GPT-4o mini, OpenAI is responding to the increasing competitive pressure in the field of AI language models. Competing companies such as Anthropic and Google already offer similar models, sometimes at cheaper rates. By combining improved performance and lower costs, OpenAI is trying to strengthen its market position while appealing to a wider range of developers and companies. The introduction of GPT-4o mini underlines the trend towards more efficient and cost-effective AI models. It remains to be seen how the new model will perform in practice and what innovative applications developers will realize with it. How do you see the development? What impact could GPT-4o mini have on your projects or the AI landscape in general?
Digital marketing enthusiast and industry professional in Digital technologies, Technology News, Mobile phones, software, gadgets with vast experience in the tech industry, I have a keen interest in technology, News breaking.
