AI has become 99 percent cheaper, putting applications in the spotlight

Using AI models has become dramatically cheaper

According to a report by the CHINA INDUSTRIAL SECURITIES CO., LTD. Research Institute of Economics and Finance, competition in artificial intelligence is increasingly about more than model performance alone. The report is based on the rapid decline in the cost of using models with comparable capabilities.

Data from Epoch AI shows that the lowest inference price for models with capabilities comparable to early GPT-4 was 37.50 US dollars per million tokens in March 2023, but had fallen to 0.18 dollars by February 2025. That represents a decline of more than 99 percent.

The Silicon Data token cost index fell from 2.0651 dollars to 0.9861 dollars per million tokens between May 28 and September 5, 2026, a decline of 52 percent. The research institute says that as open-source models improve, the bar for proprietary models is rising, meaning returns based solely on token sales could be squeezed by price competition.

Lower costs have unlocked demand before

The report uses historical examples to show that making a key resource cheaper often leads to rapid growth in demand. During the information technology revolution in the United States, the price of information-processing equipment fell by 46.5 percent between 1985 and 1990, while real consumption rose to approximately 5.5 times its original level. By 1995, prices had fallen 80.7 percent from their 1985 level, while in 2000 real consumption reached approximately 8.7 times the 1995 figure.

China's 4G era showed a similar pattern. Between 2014 and 2018, the average mobile data charge fell from 131.3 yuan to 8.5 yuan per gigabyte, a decline of 93.5 percent. During the same period, monthly data usage per user rose from approximately 0.20 gigabytes to 4.42 gigabytes, an increase of roughly 22 times.

According to the research institute, this could mean that using intelligence at a comparable level will extend beyond a limited number of use cases and appear more broadly and more frequently.

Competition could shift from models to applications

The report says that when a general-purpose technology becomes a widely accessible core capability, the focus of competition often shifts to who can use it most effectively. As computers and the internet became widespread, web portals and search engines became important areas of competition. After mobile internet was established, applications such as short-form video gained strength.

According to the CISC research institute, lower AI token prices could increase model usage and, in turn, demand for hardware. It could become easier to integrate AI into content, games, advertising, office work, transactions and local services.

This shift could also justify a reassessment of the advantages held by Chinese platform companies. In the next phase, the breadth and depth of application scenarios, the size of the user base, the strength of the surrounding industry ecosystem and engineering capabilities could become increasingly important. The report says this would allow platform and application companies with close ties to users, real-world applications and industry workflows to capture a share of the value created.

At the same time, the research institute identifies a slower-than-expected pace of AI development and commercial adoption, as well as growing global economic and policy uncertainty, as risks.