It has been revealed that one quarter of the world's highest performing generative artificial intelligence (AI) models were developed in China. As changes are detected in the AI technology hegemony that the United States previously dominated, the center of the AI industry is increasingly being reshaped into a bipolar structure between the United States and China. Other countries, except France, find it difficult even to enter the upper ranks.
According to the intelligence index of the AI performance benchmark agency Artificial Analysis on the 14th, Elon Musk's xAI-developed Grok4 ranked first with a score of 73. Following it were OpenAI's o3-pro (71 points), Google's Gemini 2.5 Pro (70 points), and o3 and o4-mini high (each 70 points), which ranked jointly in the 3rd to 5th positions. All top five models were developed by United States corporations.
The sixth place was taken by China's DeepSeek's R10528 model (68 points), while Alibaba's Qwen3, MiniMax, MoonshotAI's Kimi K2, and DeepSeek V3 were also included in the top 20 rankings. The United States maintained its technological leadership by producing a total of 14 models, including Grok, GPT, Claude, Gemini, and Llama, while China followed with a total of five models. Only France's Mistral AI's Magistral Small ranked outside the United States and China.
This indicator is regarded as the most authoritative test in the industry that comprehensively evaluates the advanced cognitive abilities of AI, which are difficult to capture with a single benchmark. Notably, the university-level MATH-500, and the science-based real-time coding assessments PsyCode and LiveCodeBench serve as standards for quantifying 'human task performance ability' beyond simple language comprehension.
China's entry into the top ranks reflects its already secured quantitative advantages in AI papers, patents, and talent. According to a report by Science released on the 11th (local time), China published 23,695 AI-related papers and 35,423 patents as of 2024, accounting for about half of the world's papers and approximately 13 times the combined total of the five countries: the United States, the United Kingdom, Canada, Japan, and Korea.
China's leading AI startup DeepSeek is expanding its footprint by achieving results comparable to American models, leveraging an open-source and low-cost high-performance strategy. HSBC, Standard Chartered, and Aramco have adopted DeepSeek models, and Amazon Web Services (AWS) is also providing DeepSeek-based models to customers.
The Belfer Center for Science and International Affairs at Harvard University analyzed in its 'Critical and Emerging Technologies Index' report published last month that 'China has an edge over the United States in AI-related data and human capital.' A joint analysis released on the 9th by UNIDO and SCMP revealed that among the world's top 100 AI experts, 50 were of Chinese nationality, half of whom were affiliated with corporations or institutions based in China.
Professor Choi Byeong-ho of Korea University's Artificial Intelligence Research Institute said, 'The overwhelming reason for the superiority of the United States and China in AI performance evaluation is that the scaling laws required for large model training have not yet been broken,' adding that 'it will be difficult for other countries to catch up with these two in terms of overall performance, and consequently, the bipolar system between the United States and China is likely to remain entrenched for the time being.' He further added, 'For other countries, including Korea, niche strategies such as specialized vertical AI or multimodal and agent-based models, or lightweight models are more realistic alternatives than competing in general-purpose AI.'