Chinese and US models

Who’s who, as of November 2026

• United States: Anthropic, OpenAI, Google DeepMind, xAI and Meta. These are the frontier labs.
• China: DeepSeek, Alibaba (Qwen), Moonshot (Kimi), Zhipu (GLM), MiniMax, ByteDance (Doubao) and Tencent (Hy).

The differences that matter at work

• Openness: the leading US models are closed, available only through APIs. Chinese labs usually publish the weights (open weights), increasingly under custom licenses with commercial conditions.
• Price: the gap is large. In November 2026, DeepSeek V4-Pro cost between $0.66 and $1.32 per million input tokens, depending on the time of day. US frontier models charged $8 to $10.
• Capability: on the Artificial Analysis index on November 12, 2026, the best US model scored 65 and the best Chinese one (Kimi K3.1 Pro) scored 46. On user-preference rankings the gap looks smaller.
• Content: Chinese regulation requires answers to follow “core socialist values”. Evaluations by NIST, a US government agency, found Chinese models echoing Chinese government narratives more often and being more vulnerable to attacks. The restrictions remain even when the model runs locally.
• Use: some US agencies and states ban DeepSeek on government devices. At the same time, US clouds such as Amazon Bedrock offer open Chinese models running outside China.

In practice

To use a Chinese model safely, the main question is where it runs. Through the lab’s API in China, your data goes there. With the weights running on your infrastructure or a cloud of your choice, it doesn’t. And test on your own tasks: a general ranking doesn’t replace your use case.