OpenAI is strengthening its presence in the competitive landscape of generative artificial intelligence (AI). Following the release of the new model 'o3,' which can conduct integrated reasoning across images, code, and web searches, attention is turning to the upcoming launch of the next-generation model 'GPT-5,' aimed at artificial general intelligence (AGI). AGI refers to an artificial intelligence capable of learning and solving various problems independently, much like a human.
On the 16th (local time), OpenAI unveiled the reasoning model 'o3,' which can recognize formulas or blurry drawings handwritten on a whiteboard or paper and solve problems step by step, as well as the lightweight model 'o4-mini,' which reduces speed and expense.
The 'o3' model is designed to step through complex tasks such as mathematics, coding, and science, while also analyzing images. Users can run Python code within the ChatGPT interface or pull web search results to refine their answers.
'o4-mini' is a smaller model specialized in simpler reasoning tasks, reducing computational load compared to 'o3.' It is characterized by improved efficiency in terms of price and speed, designed to be utilized across various device environments. The enhanced performance version, 'o4-mini-high,' has been reported to exhibit speeds surpassing human capabilities even in solving complex mathematical problems.
Scott Swingle, the founder of Avant AI and a former researcher at DeepMind, stated on his X account that 'o4-mini-high solved a newly posted Project Euler problem in 2 minutes and 55 seconds, setting a record faster than any human participant, with only 15 individuals solving this problem within 30 minutes.'
'Project Euler' is a global platform for solving mathematics and programming problems, generally appreciated for having top participants who possess excellent mathematical thinking and algorithmic abilities.
'o3' and 'o4-mini' are prioritized for users of the ChatGPT Plus, Pro, and Team subscription plans, with the high-performance 'o3-pro' set to be introduced shortly. CEO Sam Altman noted that 'we will launch GPT-5, which integrates all features, within a few months following 'o3' and 'o4-mini.'
Previously, in the roadmap announcement last February, he mentioned that 'o3 would not be an independent product but would be absorbed into GPT-5 to simplify the product line.' From GPT-5 onwards, the reasoning capabilities of the o (Omni) series and the language and tool utilization capabilities of the GPT series large language models (LLMs) will be integrated into a single engine, eliminating the selection phase for the model.
Meanwhile, OpenAI is implementing a risk-based 'Preferredness Framework' from 'o3,' systematically limiting functionalities that fail safety checks. OpenAI stated that 'GPT-5 will also follow the same protocol' and that 'as we approach AGI, safety and ethical verification will become the greatest variables in product releases.'
Thus, OpenAI's release of new models and feature expansions is directly linked to user growth. This month, the global weekly active users of ChatGPT surpassed 500 million. This marks a significant increase since ChatGPT was first released in November 2022, and the user count has risen by over 30% in just three months from 350 million at the end of last year. OpenAI has already exceeded half of its target of 1 billion users by the end of this year.
In particular, the 'Ghibli-style image transformation' feature introduced at the end of last month sparked a global meme centered on social media (SNS), igniting user growth. During the first week of its release, weekly users surged by over 150 million, prompting the need to limit daily generation due to server overload.
The continuous spike in ChatGPT's traffic has also attracted investment. Earlier this month, OpenAI raised an additional $40 billion under the leadership of SoftBank, surpassing a corporate valuation of $300 billion (approximately 425.73 trillion won). Industry views suggest that 'once GPT-5 is fully operational, the scale of graphics processing unit (GPU) clusters will increase severalfold from the current levels.'