The competition on artificial intelligence between the United States and China seems to have entered a new phase. After the “DeepSeek case” at the beginning of 2025, which had made people imagine a revolution in the sector towards cost reduction – a revolution that, however, did not happen, and indeed the focus continues to be on computing power –, now there is much talk about another Chinese company, Zhipu or Z.ai, and its latest model, called Glm-5.2.
According to the Economist, Glm-5.2 is the best-performing artificial intelligence model trained in China “to date and costs less than a tenth” of Anthropic’s Fable 5. Zhipu has promoted it as “a step forward towards frontier intelligence for everyone”, because the parameters that allow it to function have been publicly released: this essentially means it is possible to download, modify, and use it on one’s own systems without depending directly on the provider.
IS GLM-5.2 REALLY THAT IMPRESSIVE?
According to the research company Artificial Analysis, Glm-5.2 is the most open source “intelligent model on the market” and ranks fourth overall, behind OpenAi’s Gpt-5.5. Elon Musk – owner himself of an artificial intelligence company, xAi, acquired by SpaceX – wrote on X that he believes China will reach cutting-edge artificial intelligence capabilities by early 2027; according to Zhipu co-founder Tang Jie, who replied to him, “it won’t take that long.”
Andrea Stroppa, Musk’s Italian contact, also wrote on X that “Glm-5.2 (although based/copied from Claude’s models [i.e., Anthropic, ed.]) is incredibly powerful.”
Chinese companies are literally releasing impressive multimodal AI models. Seedance 2.5 is something incredible. GLM-5.2 (although based/copied from Claude’s models) is incredibly powerful. Baidu has released unlimited-OCR for automatic text extraction…
— Andrea Stroppa 🐺 Claudius Nero’s Legion 🐺 (@andst7) June 23, 2026
Comparing it with Anthropic’s Fable 5 – that is, the U.S. cutting edge in artificial intelligence, to simplify – it emerges that the latter model is 17 percent more “intelligent” than Glm-5.2 based on an average of benchmark tests. The Economist points out that in the West a comparable model to Zhipu’s was launched in February, four months ago. Furthermore, open source models – many of which are Chinese, indeed – tend to achieve better results in public benchmarks compared to private benchmarks. Since the questions used in public benchmarks to test models are known, unlike those in private benchmarks, comparing models based on different parameters might not be correct.
On this point, an analysis reported by the Economist claimed that in public tests, Chinese models lagged about four to six months behind U.S. ones; in private tests, however, the U.S. advantage was almost double, eight to ten months.
THE PROBLEM WITH CHINESE ARTIFICIAL INTELLIGENCE MODELS
The British weekly also explains how Chinese models excel in areas – such as mathematics or programming – where answers are clearly right or wrong. They have more difficulty dealing with open problems or those requiring autonomous and prolonged judgment.
This characteristic of Chinese models is linked to export controls applied by the United States against China, which prevent the latter from accessing advanced microchips. Consequently, lacking high computing power and thus unable to compete with American competition in training models, Chinese artificial intelligence companies focus on the inference phase, where the model responds to user requests. Even in inference, however, Chinese models tend to use data extracted from American systems through a process called “distillation.” Distillation, moreover, was the accusation OpenAi made against DeepSeek: OpenAi claimed that DeepSeek had improperly used the output of its models to achieve roughly comparable results but at lower costs.
ARE CHINESE MODELS REALLY MORE AFFORDABLE?
Beyond the media enthusiasm, the factor perhaps most favoring the spread of Chinese models is their price, lower than U.S. alternatives: this is why Microsoft – as reported by the New York Times – is considering introducing DeepSeek into its agent system Copilot Cowork.
But is Chinese artificial intelligence really that cheap? According to the Economist, not exactly.
It is true that DeepSeek, for example, charges only 0.8 dollars per million tokens of output for its V4 model, while Anthropic charges 50 dollars for the same service on Fable 5. But Chinese models use many more tokens to process their responses: a recent study by the Georgia Institute of Technology showed that, for the same tasks, a DeepSeek model used twenty-three times more tokens than a rival OpenAi model to achieve the same result.
Considering the significant efficiency gap, therefore, the most correct way to compare the cost of the two models is not the price per token, but the total price for all tokens used.
