According to a Stanford University assessment published on April 7, China has almost closed the performance gap with the United States in artificial intelligence (AI), with its top models reaching near parity with their American counterparts on important metrics by the end of 2024.
lead over China decreased to 0.3% on MMLU, 8.1% on MMMU, 1.6% on MATH, and 3.7% on HumanEval exams. According to the annual analysis, U.S. advantages on the same benchmarks at the end of 2023 were 17.5%, 13.5%, 24.3%, and 31.6%, respectively, which stands in stark contrast to that.
Different AI capabilities are tested by these benchmarks. HumanEval analyzes code production, MATH measures complex mathematical problem-solving, MMLU evaluates general knowledge, and MMMU checks combined text and image understanding.
Interestingly, the data revealed that the best U.S. models have typically needed a lot more training than their Chinese counterparts. According to Epoch AI statistics mentioned in the paper, the compute used for the top 10 Chinese language models has scaled roughly three times year since late 2021, which is less rapid than the global five-times-yearly rate that has been in place since 2018.
Estimated costs reflect this discrepancy; certain well-known Chinese models, such as DeepSeek-V3, have reportedly been trained for millions of dollars, while some top U.S. models have been cited for over $100 million or even close to $1 billion.
The report indicates that China is the world’s top producer of AI patents and papers. China created 15 “notable” AI models in 2024, more than Europe (3 models) and second only to the United States (40 models).
One of the best AI systems in the world, according to the research, is DeepSeek’s R1 model from China. According to the data, the performance advantage of the top U.S. model over the top Chinese model decreased from 9.26% in January 2024 to 1.70% in February 2025.
According to the Stanford study, this near-parity was achieved even though U.S.
The AI Index study, which tracks worldwide AI trends, is a comprehensive resource that is published annually by Stanford’s Institute for Human-Centered Artificial Intelligence (HAI).
