OpenAI launches new GPT-4o mini model

OpenAI announced the launch of GPT-4o mini yesterday (7/18), its most cost-effective small model to date. According to OpenAI, GPT-4o mini will significantly expand the scope of AI applications and make advanced artificial intelligence technology accessible to more people at a lower cost. GPT-4o mini’s price per million tokens is $0.24, which is more than 60% cheaper than GPT-3.5 Turbo.

Image source: OpenAI GPT-4o mini is cheaper than GPT-3.5

GPT-4o mini has powerful text and multi-modal reasoning capabilities, capable of processing text and images, and plans to add support for video and audio in the future. The model's context window length reaches 128K tokens and supports up to 16K output tokens per request. This makes GPT-4o mini ideal for applications that need to process large amounts of contextual information or require fast and immediate responses, such as customer support chatbots.

Apple devices are getting GPT-4o mini

In addition to being used in major apps and websites, GPT-4o mini will also be available on Apple devices this fall. OpenAI announced that GPT-4o mini will be integrated with Apple’s personal intelligence system Apple Intelligence to provide services for users of iPhone and other Apple devices. This move heralds that GPT-4o mini will be officially launched on the iOS 18 system, at which time users will be able to enjoy the powerful functions of this new AI model on Apple devices.

Apple released the first public beta version of the iOS 18 operating system two days ago. It is expected to be provided to iPhone Xs and subsequent models as a free software update this fall. GPT-4o mini will run on cloud servers, which means users can enjoy faster speeds and greater security without having to worry about hardware computing settings.

Apple WWDC lazy bag! How powerful is the new Siri and Apple’s artificial intelligence? Take a look at the highlights of the conference

Performance and application prospects of GPT-4o mini

The GPT-4o mini demonstrated strong performance in multiple tests, outperforming the GPT-3.5 Turbo and other smaller models. For example, in the MMLU test, GPT-4o mini scored 82.0%, while Gemini Flash scored 77.9% and Claude Haiku scored 73.8%. In mathematical reasoning and programming tasks, GPT-4o mini also achieved excellent results, scoring 87.0% in the MGSM test and 87.2% in the HumanEval test, both surpassing other competitors.

Image source: Comparison of scores between OpenAI ChatGPT-4o mini and other AI models

The launch of GPT-4o mini will bring more flexibility and lower costs to developers, allowing them to build and expand functional AI applications more efficiently. Olivier Godement, product manager of OpenAI, said that the low cost and high performance of GPT-4o mini will attract more enterprises and developers to operate with large and small models to meet the needs of different application scenarios.

Security and future development

OpenAI emphasizes that GPT-4o mini always pays attention to security issues during the development process. The model has multiple built-in safety protection measures, including filtering inappropriate content during the pre-training stage, and using reinforcement learning and human feedback in post-training to improve the accuracy and reliability of the model. These measures are designed to ensure that the model is more safe and reliable in large-scale applications. OpenAI plans to continuously monitor the use of GPT-4o mini and make improvements when new risks are discovered to further improve the security and performance of the model.

Starting today, Free, Plus, and Team users will be able to use GPT-4o mini in ChatGPT, replacing GPT-3.5; enterprise users are expected to be available next week. With the further development of AI technology, OpenAI will work to reduce costs, improve model capabilities, and make AI an indispensable part of daily life.

Image source: OpenAI ChatGPT has now replaced GPT-3.5 with GPT-4o mini