[China Shock, K-Industries on the Brink | AI] 3 China Raises Token Costs, Will Soon Hand Us the Bill

A shift in China's open-source licensing policy could deal a direct blow
South Korea must narrow the gap with the U.S. and China through frontier AI models

Chinese AI companies that entered the domestic market by cutting costs to near “dumping” levels have recently begun raising token prices one after another, showing signs of restricting access to their cutting-edge models. This has prompted calls for South Korea to strengthen its sovereign control over AI models, markets and security. The concern is that access to foreign models from the United States and China could be restricted at any time, and that without rapidly developing homegrown frontier (top-tier) models and establishing independence, South Korea will be stuck in “meaningless third place.”

According to the AI industry on November 2, Chinese companies have already begun raising token prices. DeepSeek sharply increased its application programming interface (API) fees when it announced its V4 model lineup in August. During peak hours, when usage is highest, the price of the V4 Flash model rose 4.7-fold, from $0.28 to $1.32 per 1 million tokens, while the V4 Pro model’s price increased more than 4.5-fold, from $0.87 to $3.96. Those heavily dependent on DeepSeek models have no choice but to bear the full burden of fee increases of more than four times.

Zhipu AI, which develops and offers the GLM series of models, also raised its API usage fees by up to 17% in April. Moonshot AI, whose high-performance model shook global stock markets, raised the API cost of Kimi K3 to three times that of its predecessor. Observers expect China to pursue a strategy of expanding the global reach of general-purpose models as open-weight models while restricting access to cutting-edge, high-performance models, thereby increasing reliance on its own AI models.

Starting this month, DeepSeek will charge $0.75 per 1,000 requests for calls to external tools, such as searches and gathering information from web pages. These tools were previously available for free.

Sung Nakho, Chief Technology Officer of NAVER Cloud, said, “A review of open-source licensing policies shows that they contain restrictive provisions, and there is no guarantee that current policies will remain in place in the future. This is a good strategy for a country that needs to catch up with the United States and China right away, but if China builds an ecosystem centered on its own country and then withdraws access afterward, South Korea will take a direct hit.”

Ahn Seongwon, head of the AI Policy Research Division at the Software Policy & Research Institute, said, “Many domestic developers take Chinese open-source models such as Alibaba’s Qwen, fine-tune them, and use them on in-house servers. Last month, China’s Ministry of Commerce added AI to its list of technologies subject to export bans and restrictions.” He added, “China tried to upend the U.S.-centered landscape through open source and successfully expanded and maintained its ecosystem. But now it is taking a contradictory stance by seeking direct control by the Chinese government over modifications. South Korea urgently needs a counterweight.”

As AI models become strategic national assets, countries are increasingly prioritizing their own interests in AI technology. South Korea therefore faces the crucial task of rapidly closing the model gap with the United States and China while strengthening its domestic AI models to prepare for supply risks. Choi Kyungjin, president of the Korea Association for AI and Law and a professor in the Department of Law at Gachon University, said, “Chinese AI models are currently quite attractive in terms of performance and terms of use, but their biggest problem is that their safety has not been established. Given the series of security issues involving Chinese AI models, including in the United States, it is important to fully recognize that their safety is not guaranteed. Even if Chinese AI is used, the government must establish testing to assess its safety and systemic safeguards to ensure security.”

On October 26 local time, Canadian AI company Cohere said that Alibaba’s open-weight model had attempted to send files to an Alibaba Cloud storage server in Beijing during security testing. Cohere said that, while testing “Qwen 3.8 Flash Next,” the AI model inspected the folders and network settings on the computer it was running on, then attempted to transmit a test file it had created to a server in Beijing.

Yeom Heungyeol, an emeritus professor in the Department of Information Security at Soonchunhyang University, said, “The biggest concerns with Chinese AI are its security and safety. Companies and institutions must conduct thorough safety checks to determine whether there are security vulnerabilities and whether AI models could be exploited in other attacks. At the national level, we should accelerate the development of frontier-level models so that companies and institutions can use our models, and implement policies that encourage numerous AIs to compete fiercely and advance the technology, as in the United States.”

Kim Kyungman, head of the support team at the National AI Strategy Committee, added, “To prevent technological dependence, we must concentrate all our efforts on developing frontier models. While narrowing the gap with the United States and China, we should actively develop differentiated services using large language models (LLMs) powered by AI agents.”

This content was produced with the assistance of AI translation services.