OpenAI has unveiled a new AI search and research tool, 'Deep Research,' in response to the rapid rise of the artificial intelligence (AI) 'DeepSeek' from China. This tool is characterized by overcoming the limitations of the existing ChatGPT's closed model, allowing the AI to directly explore the internet and perform research-level analysis.
◇ "Report generation after web exploration for up to 30 minutes"
OpenAI officially announced Deep Research in Tokyo, Japan, on the morning of the 3rd. Sam Altman, the Chief Executive Officer (CEO), attended the announcement, which emphasized OpenAI's search and research technological capabilities in the global AI market ahead of his visit to Korea on the 4th.
According to OpenAI, Deep Research is developed to overcome the search limitations of the existing ChatGPT, enabling the AI to explore the web based on user questions and analyze data such as documents, images, and PDFs to generate comprehensive reports at the level of research analysts. This feature will be gradually provided to paid ChatGPT subscribers and will first be applied to Pro subscribers ($200 per month) starting from that day. It will subsequently support Plus ($20 per month), Team ($25 per month), and enterprise subscribers.
OpenAI explained that Deep Research is a specialized tool for advanced researchers in finance, science, policy, and other fields that goes beyond simple searches. Unlike the existing ChatGPT model where answers are generated simultaneously with questions, Deep Research provides more precise answers after conducting in-depth searches for 5 to 30 minutes. Additionally, it generates well-documented reports with clearly stated sources, enhancing verifiability.
In particular, OpenAI emphasized its technical superiority by directly releasing performance comparison data with DeepSeek during this announcement. It recorded three times higher accuracy in the AI performance evaluation benchmark 'Humanity's last exam' compared to DeepSeek R1 and demonstrated twice the correct answer rate compared to the existing OpenAI inference model o3. OpenAI stated that it has proven overwhelming performance compared to DeepSeek R1 in various fields, including chemistry, humanities, social sciences, and mathematics.
◇ OpenAI takes steps to counter DeepSeek… strategy adjustment inevitable
Currently, DeepSeek is pressuring OpenAI with a low-cost, high-performance strategy by leveraging an open-source model. The development cost of the DeepSeek 'V3' model, released in December last year, was about 8.1 billion won, which is only 5.5% of OpenAI's GPT-4 development expense ($100 million, approximately 145 billion won). Notably, while OpenAI utilized NVIDIA's latest H100 chip, DeepSeek maintained its performance while using the H800 chip released in 2022.
Sam Altman, the CEO, said in an interview with the Nihon Keizai Shimbun on that day, 'DeepSeek is a good AI model, but it is not new.' He added, 'We will provide better models in the future,' asserting that DeepSeek's technological capabilities do not pose a threat to OpenAI.
However, industry analysts view this announcement as a response in the process of OpenAI reassessing its existing strategy following the 'DeepSeek shock.' OpenAI had previously adhered to a closed AI model, maintaining a learning approach based solely on proprietary data, but there are indications that the emergence of DeepSeek, which rapidly advanced using open-source models and 'distillation' technology, has unsettled this existing strategy. The distillation technology is a method that compresses and transfers the knowledge learned by large AI models to smaller models, allowing for high performance at lower expenses.
OpenAI is currently in a situation where it is falling behind DeepSeek in price competition. In response, OpenAI has shown moves to secure competitiveness by launching a new free AI model called 'o3 Mini' on the 31st of last month.
Professor Kang Jae-woo of Korea University stated, 'Deep Research is a model designed to perform complex multi-layered inference, developed to enable more sophisticated analysis by combining internet search capabilities.' He added, 'This had been the direction the industry has been researching, and it seems that OpenAI adjusted the announcement timing in response to the emergence of DeepSeek.'
He further noted that 'the fact that DeepSeek learns high-performance AI at a low expense while being fully open (MIT license) means that anyone can utilize it, shaking up the ecosystem.'