In China there is what is popularly known as the “Great Firewall”a large security wall that prevents its citizens from accessing certain services. At the same time, there are foreign companies that block their services in China, this is what is happening with AI tools like ChatGPTbut the law is made, the trap is made.
Gray market. They tell it in South China Morning Postthere is a whole flourishing market of services that promise to provide access to American AI models, such as Claude or Gemini, avoiding the restrictions imposed on both sides of their borders. On online sales platforms such as Taobao or Xianyu, unlimited subscriptions to Claude Code, Gemini and ChatGPT are sold with low latency and without VPN.
These platforms have become a solution for Chinese developers who want to access American models to program, debug or use multimedia generation services.
How they do it. Access is done through what is known as ‘shadow APIs’ which, in essence, is an intermediary. What they do is set up proxy servers outside of China and divert all user requests there, so that those external servers are the ones that actually call the official APIs of models like Claude or Gemini and then return the “masked” response as if it were a local service, without the need for a VPN or foreign payment methods.
It pays for them. According to the developers cited in the South China Morning Post piece, they resort to these ‘shadow APIs’ because they simply consider them to be tools that are clearly superior to the local offering. This translates into more precise code, fewer hallucinations and less time correcting bugs than with Chinese models which, according to what they say, still invent functionalities or fail more often. In addition, these services give them almost complete access to models like Claude Opus or Gemini, with huge context windows (up to a million tokens), without having to fight with VPNs or foreign payment methods.
Wiles. All that glitters is not gold and there are also advertisements that do not fulfill what they promise. Some of these services advertise full access to models like Claude, but are actually processing requests with cheaper Chinese models like Qwen or MiniMax. Additionally, there is the risk to privacy as all traffic goes through an anonymous intermediary who could do whatever they wanted with often sensitive data.
Frontier Model Forum. Is a coalition formed by several AI companies dedicated to the security and regulation of border models, but in practice it is functioning more as an intellectual property defense mechanism. Recently OpenAI, Anthropic and Google announced that they were working together to curb copying of their models by coordinating the sharing of suspicious usage patterns and distillation attack detection techniques through this common forum.
Image | Xataka
GIPHY App Key not set. Please check settings