After starting research, it frequently occurs that the model relies on its own database and is unable to recognize the latest information. For instance, when I asked it to compare the performance of Gemini 2.5 flash and GPT 4.1, DeepResearch said GPT 4.1 had not been released and then chose GPT4o as the comparison subject. During this process, it often substituted ‘o4mini’ with ‘4omini’, making the research report unusable.
Even after subsequently searching and finding that the GPT4.1/o4mini models exist, it still proceeded with the research planned earlier focusing on GPT4o/4omini.
I have encountered this situation many times. Sometimes it will even research Trump’s tariff policies from his first term instead of the policies from his second term, even after searching and finding that Trump began his second term in 2025