“OpenAI has launched the deep reasoning model ‘DeepResearch’ in response to the challenge posed by Chinese AI startup DeepSeek. This model achieves approximately three times higher accuracy in AI performance evaluations compared to DeepSeek’s latest reasoning model ‘R1’ and boasts about twice the accuracy of OpenAI’s previously best-performing model ‘O3’.
OpenAI made a surprise announcement of ‘DeepResearch’ during a live-streamed event in Tokyo, Japan, on the 2nd (local time). This model is an enhanced version of the existing ‘O3’ reasoning model, now featuring internet search capabilities. It analyzes and synthesizes vast amounts of online materials to generate more in-depth responses. OpenAI explained, ‘ChatGPT explores hundreds of online sources to create comprehensive reports at a researcher level,’ adding that it can complete tasks in mere tens of minutes that would take a human several hours.
DeepResearch goes beyond simple information provision to support in-depth analysis for experts in finance, science, policy, and more. Unlike the existing ChatGPT, which generates answers immediately upon questioning, DeepResearch undergoes a searching and analysis process lasting 5 to 30 minutes to produce more precise results. This is similar to a feature Google introduced as a pilot last year.
OpenAI emphasized, ‘All output results include clear citations and summaries of the thought process, making reference and verification easy. It is particularly effective for non-intuitive information searches that require direct exploration of multiple websites.’
DeepResearch recorded a 25.3% accuracy rate on the ‘Humanity’s Last Exam,’ one of the most challenging benchmarks for evaluating AI performance. This significantly surpasses the 3.3% accuracy of GPT-4o, as well as the 9.1% and 9.4% of OpenAI’s previous reasoning model ‘O1’ and DeepSeek R1, respectively. Compared to the 13.0% accuracy of OpenAI’s latest reasoning model ‘O3 mini’ when using high-performance resources, DeepResearch’s performance is exceptional.
OpenAI stated that ‘over 3,000 multiple-choice and short-answer questions were tested across more than 100 subjects, including linguistics, rocket science, classical literature, and ecology. Particularly in chemistry, humanities, social sciences, and mathematics, there were high achievements, and it demonstrated a human-like approach to exploring specialized information.’
Initially, DeepResearch is available to ChatGPT Pro users (priced at $200 per month), with a limit of 100 questions per month. The service target is planned to expand to include Plus and Business users. OpenAI mentioned, ‘DeepResearch is a compute-intensive model requiring substantial computing resources,’ and announced plans to develop a cost-effective version that can provide high-quality results with smaller models in the future.”