[Big read] China’s AI catch-up is forcing Silicon Valley to cut prices

(Edited and refined by Grace Chong, with the assistance of AI translation.)

Los Angeles-based Chinese entrepreneur Li Jianian deployed Chinese company Zhipu AI’s GLM-5.2 on a local server, using it for some of the tasks he had used AI coding tool Claude Code to do. From his experience, Zhipu AI’s performance is “already on a par with Claude” for around 70% of his day-to-day work. “There’s no need to use the most expensive model for everything,” he added.

Halfway around the world, in Chengdu, Sichuan, video creator Cheng Xiao (pseudonym) is facing a different problem: his Claude account has been blocked again.

Cheng switches between AI models, but still considers Claude “smarter” at complex coding. The problem is access. Anthropic, Claude’s parent company, does not offer its services in China, forcing Cheng to keep his computer connected to a VPN (virtual private network) and rely on an intermediary to top up his account. When his latest account was blocked, he switched to OpenAI’s coding tool, Codex.

Amid intensifying China-US tech rivalry, users in both countries are putting the other side’s models into their own toolkits. Some are looking for better value for money, while others need the more powerful option.

At present, Claude’s Fable 5 remains the top-ranked model on several leaderboards. But Zhipu AI’s GLM-5.2, Moonshot AI’s Kimi K3 and Alibaba’s Qwen3.8-Max have all been released in recent months, with their performance on multiple tasks approaching that of Claude, while their prices are also lower.

Zhipu AI’s GLM-5.2 costs US$4.40 per million output tokens, compared with US$10 for Claude Sonnet 5.

According to OpenRouter statistics, two of the ten large language models with the highest token processing volumes over the past month as of 14 August were from the US, while the other eight were from China. However, while Chinese models account for higher usage volumes, US models still attract more spending. Across nearly 30 task categories, Chinese models accounted for the largest share of spending in only seven, while leading US models such as ChatGPT and Claude dominated most of the rest.

Kendra Schaefer, head of digital research at Trivium China, said in an interview that it is common for companies and users to mix and match different models. Users place different priorities on performance, cost and data security. “There’s not one market for AI models; there’s multiple market segments,” she noted.

She said that the US currently lacks a good competitive offering in the “controllable and cheap market segment”, which is an important reason why China’s open-weight models have been able to enter the toolchains used by US developers.

Deepening regulatory boundaries

But as the regulatory boundaries between China and the US become increasingly clear, it remains uncertain how long users will be able to continue calling on models from both sides as needed.

US users are concerned that growing attention to Chinese models could prompt Washington to impose restrictions on their use. The US has already banned the use of DeepSeek models on government devices at some agencies, citing security concerns.

US Treasury Secretary Scott Bessent said publicly in July that the US government would investigate whether Chinese open-weight models were involved in intellectual property theft. If infringement were confirmed, he said, the US could consider financial sanctions and adding the companies to the Entity List.

China, too, has begun debating whether to continue making advanced models freely available without restrictions. According to Reuters, Beijing is considering restricting overseas access to advanced AI models. Chinese state media outlet Yuyuantantian (《玉渊谭天》) also said in a post that “supporting AI open models does not mean that all capabilities should be disseminated unconditionally”.

Cheng said that even if his appeal succeeds and he regains access to his Claude account, he would not rule out the possibility that Chinese authorities could one day proactively block its use. In July, China’s Ministry of Industry and Information Technology warned that the Claude coding software contained built-in monitoring that could transmit sensitive information, posing a serious security “backdoor” risk.

Li joked, “Who knows? We might end up having to ‘bypass the Great Firewall’ ourselves to download the latest Chinese models.”

Such concerns may still be premature, but the signs increasingly suggest that China-US AI competition is entering a new phase.

China: a latecomer playing catch-up

In the past, US tech giants invested heavily in computing power in pursuit of artificial general intelligence (AGI), while China, facing computing constraints, focused on low costs, open weights and practical applications. China was widely seen as the latecomer playing catch-up.

But the rapid progress of Chinese AI models has forced the US to reckon with a different possibility: even if China has yet to overtake the US at the frontier, its low-cost, open-weight models and widespread diffusion could still undermine the US’s existing technological and ecosystem advantages.

Kyle Chan, a fellow at the John L. Thornton China Center at Brookings Institution, took part as an expert witness in a US House Select Committee on the Chinese Communist Party hearing on China-US AI competition in April.

He told Lianhe Zaobao that Washington has two concerns about the narrowing gap between US and Chinese models. First, if Chinese models with comparable capabilities become widely available, they could circumvent the US AI regulatory framework. Second, powerful open-source Chinese models could disrupt the business models of companies such as OpenAI and Anthropic, making it harder for them to justify the enormous investments they have made.

Professor Lin William Cong, President’s Chair Professor at Nanyang Technological University (NTU) and an associate dean of Nanyang Business School, went further, pointing out an inherent weakness in the commercial moat of large language models: they are products that continually “teach” their competitors. Even when model weights are not made public, large volumes of model outputs can become material for other models to learn from.

This is precisely what US companies and officials have repeatedly accused Chinese companies of doing through “distillation”. Cong said that if model capabilities can be transferred through distillation, the technological lead underpinning high valuations and massive investment may be far less secure than imagined.

Different starting points

Chan said the China-US AI race is proving more complex and messier than expected. Washington is beginning to recognise the importance of widespread technology diffusion and adoption, as more voices call for the US to develop strong open-source models to compete with China.

Last year, Washington began promoting a plan to export a “full-stack American AI technology package”, bundling a range of AI products including AI models, chips, servers and data centres for export to allies. The initiative is seen as a US response to China’s efforts in recent years to export its tech model to the Global South.

This year, the Donald Trump administration has sought to pressure allies to choose between the AI ecosystems led by China and the US, narrowing opportunities for China to engage in supply chain and technological cooperation with them.

Reuters reported on 15 August that the US plans to write to dozens of members of the AI alliance behind the Pax Silica initiative, warning them against joining the Beijing-led World Artificial Intelligence Cooperation Organization. Those who do so would be excluded from Pax Silica. The warning was that “to be part of everything is to be part of nothing”. Pax Silica focuses on cooperation in AI supply chains.

Yasir Atalan, deputy director and data fellow in the Futures Lab at the Center for Strategic and International Studies, said in an interview that Washington could step up export controls on Chinese models in future. For now, however, its focus is on establishing mechanisms to promote the export of US AI technology.

Companies, too, are stepping up efforts on open-source models and pricing. OpenAI announced at the end of July that it would cut the prices of GPT-5.6 Luna and Terra by 80% and 20%, respectively. More recently, Anthropic also announced that it was abandoning plans to raise the price of Sonnet 5 and would permanently retain its introductory promotional price. Meta and Nvidia have also released new open-source models.

China and the US started from different positions in AI development. The US invested heavily in computing power in pursuit of superintelligence, while China, constrained by limited computing resources, sought to compete through open models and lower costs. But their goals are increasingly converging: both want to develop the most powerful models at the top of the pyramid while reaching a broader range of markets.

The bigger goal beyond tech diffusion and adoption

In official statements, the Chinese government rarely mentions AGI. But that does not mean China has no intention of competing at the frontier of AI. DeepSeek founder Liang Wenfeng has even stated bluntly that achieving AGI is the company’s true goal.

NTU’s Cong said China’s leading AI companies are scaling to the limits of available hardware. Kimi K3, for example, has 2.8 trillion parameters. “This is not the behaviour of an ecosystem content with cheap application diffusion,” he noted. In his view, laboratories are pursuing frontier capabilities, while the government places greater emphasis on technology diffusion and industrial applications — both are true simultaneously. 

Tan Poh Hwee, president of the Asia Academy of Digital Economics, believes China’s AI strategy has two layers. The first is to drive AI adoption by integrating it into real-world production, turning it into a new engine of economic growth. The second is not to give up competing at the frontier while expanding influence through open-source large language models.

But further progress at the frontier remains constrained by computing power, while the low-cost advantage enjoyed by some Chinese models is also coming under pressure. As business expansion drives up computing service fees, Zhipu AI has raised its prices several times this year. After launching Kimi K3, Moonshot AI also temporarily suspended new subscriptions when demand exceeded its computing capacity.

‘Learning to fight through fighting’

Facing constraints, Chinese companies are improving training efficiency and expanding applications while steadily narrowing the gap with the most advanced models. Tan describes this as “choice made out of necessity” as well as “learning to fight through fighting”, something that Chinese people are good at.

He pointed out that model capabilities are closely tied to a country’s future national power. In fields such as scientific research and the military, the most powerful models are still necessary; otherwise, China risks developing a strategic capability gap with the US. “Just like nuclear weapons, both sides must possess them at the same time to establish a balance of power,” he said.

Atalan holds a similar view. He says that “good enough” models can handle a wide range of commercial tasks, with the key being to integrate them into business operations. The most powerful models, meanwhile, can help achieve breakthroughs in highly challenging areas such as science and mathematics, which in turn can affect economic and military capabilities.

He asserted, “Whoever leads in the frontier, will likely sell its AI stack to the globe more.”

The differing demands for model capabilities are pushing the China-US AI competition to different levels. In most commercial and everyday applications, “good enough” open-source models will continue to cross borders. Frontier models, however, will become protected strategic assets.

This is also a competition that will be difficult to stop. Cong pointed out that the boundaries between these tiers are not fixed. Capabilities that are considered cutting-edge today may soon become basic tools accessible to ordinary businesses as open-source models catch up, costs fall and the technology spreads. New strategic boundaries will continue to move forward.

Commercial development vs government control

Competition between China and the US, the world’s two leading AI powers, is driving the development of more powerful and potentially more dangerous models. Public reports show that since July, there have been at least seven cases in which AI models found vulnerabilities during sandbox testing, “jailbroke” their way out and breached real-world systems.

International calls for the establishment of AI safety rules are growing louder. China and the US are also planning to hold an official AI dialogue in September to discuss the safety risks of advanced AI models. However, it remains uncertain how much information the two sides will be willing to share, and how much common ground they can reach, amid the current competitive environment.

Matt Sheehan, a senior fellow at the Carnegie Endowment for International Peace who has long followed global technological change, wrote that China and the US will not move in lockstep on AI safety, and that it is currently difficult for them to forge mutually binding agreements with real teeth. The two countries can, however, improve the safety of their respective models based on their own security interests, and then engage in a minimum level of communication.

He said that if each side understood what safety measures the other had put in place, it could reduce the worst-case assumption, born of a lack of information, that “the other side is sacrificing any action on safety in a reckless race for building the most powerful system”.

Cong pointed out that China and the US will ultimately face the same governance challenge: how to prevent the uncontrolled diffusion of the most dangerous frontier capabilities.

Once frontier-level weights are widely distributed, they cannot be recalled and could be used to accelerate high-risk activities such as cyberattacks and biological design. As a result, this creates a wedge between a firm’s commercial incentive to release the most advanced models and the government’s goal of controlling the spread of sensitive capabilities.

Cong said that if AI approaches superintelligence, that wedge will become particularly pronounced. In such circumstances, “it is one of the few areas where US and Chinese interests in verification and dialogue genuinely overlap”.

Atalan, meanwhile, warns that if China and the US each integrate increasingly powerful AI into their military systems without a common framework and confidence-building mechanisms, it could alter their existing strategic calculations. Once either side begins allowing AI to participate in critical military decision making, the other side will be forced to adjust its deterrence logic, increasing the risk of miscalculation and strategic instability.

Open vs closed-source

However, even if both China and the US recognise the risks posed by the proliferation of frontier capabilities, there is still no simple answer as to whether those risks should be reduced through “open source” or “closed source”.

A recent case has changed some people’s views. An OpenAI model infiltrated the production database of Hugging Face, a US AI community, to search for test answers. When attempts to trace the model’s origins using leading Western large language models were unsuccessful, Hugging Face ultimately used Zhipu AI’s GLM-5.2 to contain the attack.

The Brookings Institution’s Chan said attitudes towards open-source models are complex within the US AI safety community. Some worry that locally deployed models are difficult to control, while others believe they can strengthen the defensive capabilities of those under attack. “So far, it seems like the Trump administration is leaning towards the view that open-source models are a net asset for AI safety and have excluded them from the latest regulatory framework,” he observed.

This article was first published in Lianhe Zaobao as “中国缩小与美国前沿差距引恐慌 华盛顿祭出防堵大招 硅谷AI罕见降价应战”.

Related: Can AI’s sky-high valuations survive disclosure? | Mythos vs DeepSeek: Why cheap AI could tip the US-China tech race