Research by Zilan Qian from the Oxford China Policy Lab has revealed a large grey market for API tokens in China, where access to leading models, such as Anthropic Claude, is sold at discounts of up to 90%. Intermediaries use complex model-swapping schemes and data collection methods to illegally train other neural networks.
What Happened
A massive system for reselling API tokens through proxy services has been deployed in China. The scheme includes using hired individuals in other countries to pass identity verification, subscription arbitrage, and a "model swapping" method, where requests to top-tier models are redirected to cheaper and less capable architectures. Additionally, intermediaries engage in data distillation, collecting user queries and reasoning chains to train third-party LLMs.
Context
This infrastructure was created to bypass regulatory and technical barriers established by Western AI companies. The extreme 90% discounts are subsidized by exploiting user data and stealing intellectual property through illegal model distillation.
Why It Matters for the Industry
Grey markets pose a direct threat to intellectual property and accelerate the illegal distribution of data. This forces developers like Anthropic and OpenAI to tighten geoblocking systems, account verification methods, and implement model authenticity verification tools. In the long term, this could lead to a decline in industry-wide model quality due to model collapse cycles caused by mass training on data obtained from such networks.
Why It Matters for Users
For users, using such APIs carries critical risks: immediate leakage of confidential corporate data and receiving low-quality results due to the substitution of strong models with weak ones. Saving 10x can lead to a loss of information control or making incorrect decisions based on inaccurate neural network responses.
Sources
Author
Look at AI, Editorial Team