Artificial intelligence Photo: VCG
Chinese AI developers have recently launched “migration” services to support their customers to seamlessly switch from Claude AI models to indigenous AI tools that match performance but cost less.
Earlier this month, a Bloomberg report said that American startup Anthropic is to bar Chinese-run entities from using its AI services.
The US wields the big stick of blockades, while China adopts an open mindset to embrace the market, presenting two varied logics of technology governance, industry experts say.
From a commercial perspective, this supply cutoff by Anthropic would provide huge market opportunities for China’s own large model vendors, Liu Wei, director of the human-machine interaction and cognitive engineering laboratory with the Beijing University of Posts and Telecommunications, told the Global Times.
Last year, when OpenAI cut off its API services to China, the move triggered a wave of “migration” enthusiasm among Chinese AI companies. This time with Claude’s cutoff, domestic vendors are evidently better prepared, with faster response speed, Liu said.
One of China’s leading large model companies, Zhipu AI, launched a “migration” plan on September 5, immediately following Anthropic’s cutoff announcement, claiming full compatibility with the Claude protocol. Users only need to replace the API URL and key to seamlessly switch to the Zhipu-developed GLM model API, and new users are provided with 20 million tokens of free trial quota.
According to the global SWE-reBench leaderboard, a global ranking list that reflects the performance of various models, Zhipu AI’s GLM-4.5 model’s programming capability is leading in global ranking, with its performance close to Claude Sonnet 4.
“Large models form the foundation of AI technology infrastructure, and large models competition is also technology strategy competition. China should have full-stack autonomous innovative large model products. Therefore, Zhipu’s entire technology stack—from underlying algorithms and pre-training frameworks to team building and domestic hardware adaptation—is independently developed by our own team, ensuring full autonomy and controllability,” Zhipu AI’s CEO Zhang Peng told Global Times.
Another leading AI company, Moonshot AI, rolled out its updated Kimi K2-0905 model on the day of Anthropic’s cutoff. The new version significantly improves performance in real-world programming tasks. The company provides a high-speed version API, supporting output speeds up to 60-100 tokens/s, while being compatible with both OpenAI and Anthropic API formats.
SenseTime, an established AI player, has moved quickly, offering a 50-million-token trial package for users migrating from Claude to its SenseNova platform. The company has assigned dedicated “migration” consultants to provide full training and support throughout the transition.
A source close to Moonshot AI told Global Times that even without external “bans” or “cutoffs,” users would choose more suitable solutions.
In the SWE-bench Verified that more reliably evaluates AI models’ ability to solve real-world software issues, the Kimi K2-0905 open-source model has already reached up to 95 percent of Claude Sonnet 4’s performance, but at less than 1/5 the price, making it a good alternative on the market. The company will use this momentum to accelerate independent R&D iterations and provide developers with more complete alternative solutions in an open manner.
Shen Yang, a professor at Tsinghua University’s School of Journalism and Communication and the College of AI, described the US blockade as mere “venting of strategic fear” with minimal practical impact on China.
China’s swift response underscores the resilience and agility of many Chinese AI players. This episode highlights a fundamental divergence in AI ecosystem philosophies: the US prioritizes “technology weaponization,” while China champions “technology popularization,” Shen noted.
“This is not the first time American AI firms have restricted services provided to China. In June 2024, OpenAI decided to stop providing API services to China. Anthropic’s action this month is seen as another escalation of US technical blockade against China,” said Liu.
He believes that certain US companies are pushing for technological decoupling, which reflects deep anxiety in the US AI community about China’s rapid AI development.
Industry experts point out that this supply cutoff incident once again highlights the importance of autonomous and controllable technology. From semiconductors to AI large models, China’s tech industry is undergoing a forced breakthrough process. Each external blockade may bring short-term pain, but it accelerates the development of alternative technologies in China.
Shen emphasized that the US containment of China will produce limited effect. “Large models differ from chips: chip production links are highly concentrated in the hands of a few companies, while AI large models are replicable products of algorithms and data. In other words, chip blockades may create years-long supply vacuum, but large model blockades are more psychological and ecosystem-level suppression, hard to create absolute barriers,” Liu said.
“If in 2023 there were still people doubting that Claude and GPT-like models were irreplaceable, then by 2025, China’s domestic large models are close enough in performance, with migrations almost seamless,” Shen said.
However, the scholar cautioned that whether domestic large models can truly replace needs market validation. “In some aspects, there may still be about 20 percent gap between the two, but this 20 percent is precisely the core breakthrough point for domestic vendors to complete in the next few years.”
Currently, to unilaterally maintain technological advantages and prevent other countries from borrowing their results, a majority of US large AI models would adopt a closed-source development route. In July, the New York Times reported that Meta—long an advocate of open-source sharing—plans to pivot toward closed-source AI models as well.
In contrast, Chinese AI large models have consistently embraced open-source development philosophy, continuously achieving technological breakthroughs.
Technology news website Tech Wire Asia said in an article in July that less than a year later, US’ strategy of protecting American AI innovation from Chinese competition has backfired. “China’s open-source AI models now dominate global rankings, with Chinese companies DeepSeek and Alibaba delivering world-class performance that rivals – and in some cases surpasses – that of their Silicon Valley peers,” the website said.
“The question now is not whether open-source AI will challenge proprietary models, but whether the West can compete with China’s collaborative, accessible approach to AI development. The future of AI may indeed be open – and increasingly shaped by Chinese innovation,” it concluded.