Interest in Chinese AI companies is again drawing focus in the global tech industry after cooling somewhat following the DeepSeek shock in January 2025.

More companies are feeling burdened by the cost of using the latest models from U.S. firms such as OpenAI and Anthropic across their organisations. Chinese AI models positioned as offering value for money are emerging as an alternative.

Perceptions of Chinese AI models have moved well beyond being cheap and usable at a basic level. Some benchmarks are being released showing results that outperform OpenAI and Anthropic.

Z.ai, a Chinese AI startup known as Zhipu, recently announced GLM 5.2, a high-performance open-source model. GLM 5.2 is being assessed as having the best performance among AI models developed in China so far. Its price is also about one-tenth that of Anthropic Fable5.

Moonshot AI also announced the latest version of its "Kimi" in June and presented benchmark results comparable to top products from Anthropic and OpenAI.

As of June 24, local time, research firm Artificial Analysis ranks GLM5.2 as the most intelligent open-source model on the market. It placed third among all models, behind Anthropic Fable5 and OpenAI GPT 5.5. It is ahead of Google Gemini 3.1 Pro Preview.

After Z.ai released GLM 5.2, Elon Musk, Tesla's CEO, projected on social media platform X (Twitter) that China would reach the level of today's advanced frontier models by early next year. Tang Jie (탕 제), a co-founder of Z.ai, countered that "it will not take that long."

The Economist reported that, at the current point in time, averaging the results of several benchmark tests shows Fable5 is about 17 percent more capable than GLM 5.2. A Western model with performance similar to GLM 5.2 was released about 4 months ago. Chinese models often post strong results in areas with clear right and wrong answers such as maths or coding, but tend to make mistakes on open-ended questions or tasks requiring sustained independent judgment, The Economist said.

Chinese models are also criticised for frequent service disruptions due to a lack of computing power, or for slowing down during periods of surging traffic.

On price, Chinese AI models appear competitive. DeepSeek charges $0.87 per 1 million output tokens for its newly released v4 model, far cheaper than Anthropic Fable5 at $50. This is also why the number of U.S. companies using DeepSeek has been rising recently. Customer data from corporate payments platform Ramp shows a surge in DeepSeek service payments in the United States, and Microsoft is also reported to be reviewing the use of DeepSeek models in its flagship Copilot AI chatbot.

On OpenRouter, an AI model marketplace, 6 of the 10 most popular models were developed by Chinese companies including DeepSeek, Tencent and Xiaomi.

The New York Times said OpenRouter CEO Alex Atallah (Alex Atallah) said, "Cost issues are driving adoption of Chinese models." Many Chinese models are open source and can be used for free. As more companies use Chinese AI models, valuations of Chinese AI companies are also soaring. Z.ai, listed in Hong Kong, surpassed 1 trillion Hong Kong dollars ($128 billion) after the release of GLM 5.2.

The South China Morning Post said a JPMorgan report drove Z.ai's stock rise. JPMorgan last week raised its 2026 to 2030 revenue forecast for Z.ai by 7 to 16 percent to reflect the launch of the GLM-5.2 model. JPMorgan forecast Z.ai's 2026 revenue would surge by more than 534 percent and expected it to turn profitable in 2028.

DeepSeek is also reported to have raised a $7.4 billion investment at a valuation of more than $50 billion.

Some say the assumption that Chinese AI models are cheap may be wrong. According to a study updated in June by Du Zheng (두정) of the Georgia Institute of Technology and co-authors, when performing the same task, DeepSeek models used 23 times more tokens than a competing OpenAI model to produce essentially the same results. Because this efficiency gap is very large, models should be evaluated based not on price per token but on the total cost of tokens actually used, The Economist reported.