New test results from Artificial Analysis show that China's open weighted AI model still lags behind its American counterpart.

Business Insider reported this week that Artificial Analysis, an independent company specializing in measuring AI performance, improved its AI model testing system to better reflect how people actually use the technology. One important change is that many tests are now kept secret, making it difficult for AI companies to train models on them.

The new assessment results are significantly different, showing that US companies are further ahead than the old estimate, especially with complex scientific tasks. The best open weight model is 12 points behind the leading model, a sharp increase compared to the gap of 3 points under the old system. Specifically, while the top model Claude Opus 5.5 of Anthropic (USA) scored 58 points, Kimi K3.1 Pro of startup Moonshot AI (China) scored 46 points.

China releases many powerful AI models that other businesses can download and adapt, often at a lower cost than American alternatives. This makes American experts worry that companies like OpenAI or Anthropic will find it difficult to charge high fees for their proprietary technology. However, the new rankings suggest this risk may have been exaggerated.

Sridhar Ramaswamy, CEO of cloud computing company Snowflake, shared the new results on social network

According to Business Insider, Ramaswamy said that when AI companies know exactly how the model will be tested, they can train it to excel at those tasks. However, this does not mean the model works equally well in real life.

Even proponents of the open-weight model acknowledge the uncertainties. "I think the old rankings rated the open weight model too highly," said Nathan Lambert, co-founder of Trillium Labs, a large language modeling research company.

In addition to standard tests, another factor to evaluate AI models is actual usage. Last week, Dax Raad, co-founder of software development company OpenCode, looked at model traffic via OpenRouter, a popular platform that allows programmers to access AI models from a variety of providers.

He estimated that almost all of DeepSeek's traffic via OpenRouter on October 27 could be handled by just one high-performance Nvidia server cabinet, at a cost of about 6-7 million USD. This computing power is surprisingly low considering the huge attention paid to Chinese AI models today. However, some experts note that OpenRouter only represents a portion of the market, so this data does not fully reflect the demand for an open weighting model.

Data from the Hugging Face platform shows that the Chinese model accounts for nearly half of all open source AI downloads from February 2025 to 2026. Cien Solon, CEO and founder of the LaunchLemonade agent platform, said in July that Claude and ChatGPT still dominate in terms of usage, but GLM 5.2 of Chinese startup Zhipu AI launched in June is also in the top 5 of this platform.

"The Chinese model is especially attractive to US businesses as AI costs skyrocket. Previously, they prioritized AI applications regardless of the model, but now they are starting to pay more attention to costs," Kyle Chan, an expert at the John L. Thornton China Center of the Brookings research organization, explained to CNBC in July.

Chan then commented that, although often "a fraction of the cost", the performance of Chinese AI models is increasingly improving, even "close to the leading advanced US models". He estimates they are only 6-9 months behind.

According to Lambert, the open weighting model remains highly competitive in tasks such as software development, while also giving businesses greater control at a low cost. However, new evidence challenges the notion that Chinese AI companies are quickly catching up with leading US labs and threatening their business models.

Thu Thao summary