When the people asked the latest artificial intelligence (AI) model to solve the common Japanese university entrance test question, ChatGPT recorded a higher percentage of correct answers.
The Nihon Keizai Shimbun reported on the 20th that ChatGPT’s correct answer rate was 97 percent, ahead of Gemini and Claude (91% each), as a result of such experiments on OpenAI’s “GPT 5.2 Thinking,” Google’s “Gemini 3.0 Pro,” and Antropic’s “Claude Opus 4.5” with Japan’s AI start-up Lifeprompt. ChatGPT received perfect scores in nine subjects, including Mathematics 1 A, Mathematics 2 BC, and Chemistry, in this experiment, which was conducted on the 17th and 18th by solving 15 major questions in the common college entrance test within a limited time. Although the percentage of correct answers in Korean (Japanese and 90%), comprehensive geography and exploration (91%) was somewhat low, Japanese history (97%) and English (96.5%) were close to perfect, and the overall average correct answer rate was 97 percent.
The Generative AI model of OpenAI was evaluated to be increasing in precision as the number of correct answers for the common Japanese university entrance test gradually rose from 66% in 2024 and 91% in 2025.
Gemini struggled in physics (81%) and geography comprehensive and exploration (82%), and Claude struggled in geography comprehensive and exploration (72%), showing an average correct answer rate of 91%.
The percentage of correct answers in these three AI models was significantly higher than the average 60% of test takers expected by the Japanese cram school Kawai Academy.
However, Gemini and Claude completed one subject in the shortest four minutes, and math and Japanese also filled out the correct answer in up to 20 minutes, while ChatGPT took two or three times longer than this, Nikkei said.
The three AI models showed strength in science subjects and struggled in Japanese and geography. Geography in particular revealed weaknesses in color classification and solving problems by combining graphs and maps of 47 local governments.
“If you use an open AI model for the use of finance and legal work that requires high precision, and use Google and Antropic models for tasks where speed is important, you can maximize AI’s strengths,” said Satoshi Endo, CEO of Lifeprompt.
As Gemini 3.0 recently shook the AI market with favorable reviews, competition for AI technology is intensifying, with OpenAI launching a counterattack by early deployment of GPT 5.2.
JENNIFER KIM
US ASIA JOURNAL
