The Truth Behind "Claude 90% Off": A Channel for Stealing AI Training Data

On China's black market, Claude is being sold at around 10% of the official price, and the problem is not simple resale — it's a structure that collects users' prompts, model responses, and reasoning process (Chain of Thought, CoT) to resell as training data.

Zhilan Chen of the Oxford China Policy Institute described the so-called API Proxy Economy, noting that proxy networks operate openly on GitHub, Telegram, Taobao, and elsewhere. The methods used to lower prices include:

• Mass-creating free trial accounts to resell API access
• Signing up for paid plans with stolen credit cards to distribute access
• Splitting a Max plan costing around $200 per month among multiple users
• Model swapping, where users believe they're using Opus but responses are actually routed through the cheaper Haiku or open-source models

Researchers at Germany's CISPA Helmholtz Center for Information Security analyzed 17 proxy services and found that on a medical benchmark, the official API scored 84% accuracy while the proxies scored only 37%. This means the cheaper the relay service, the greater the quality degradation.

The bigger risk is that operators store prompts, answers, and reasoning processes to sell as datasets. Anthropic revealed that in February 2026, over 16,000,000 queries came from about 24,000 fraudulent accounts.

Organizations using coding agents and internal tools need to be especially careful. Going through unverified proxies can leak source code, API structures, and even authentication credentials to external servers, making it necessary to re-examine AI usage paths and traffic.