Anthropic Blocks Chinese Entities After Grey Market Sells Claude API at 90% Off

Anthropic has identified a large-scale security breach involving its Claude API, where grey-market proxy services in China resell access at steep discounts. These unauthorized networks operate through platforms such as GitHub, Taobao, and Telegram, offering prices up to 90% lower than official rates. The company blocked Chinese-controlled entities from accessing the Claude API in September 2024 in response to these activities.

CISPA audit reveals users requesting premium models often receive cheaper alternatives

The proxy networks, referred to as transfer stations, sustain their low prices through illicit methods including stolen credentials and model substitution. Upstream operators farm free API credits, exploit corporate discounts, or subdivide $200 Max subscriptions to generate supply. Some accounts are purchased using stolen credit card details, while others recruit individuals in lower-income countries to bypass photo ID and selfie verification requirements.

Audits by the CISPA Helmholtz Center revealed that users requesting premium models like Claude Opus often receive responses from cheaper alternatives such as Sonnet, Haiku, or Qwen. One proxy-marketed Gemini-2.5 service scored only 37% on a medical benchmark compared to nearly 84% for the official API. Operators harvest all user prompts, responses, and reasoning chains to resell as training data for AI distillation campaigns.

Anthropic identified roughly 24,000 fraudulent accounts linked to Chinese labs including DeepSeek, Moonshot AI, and MiniMax in February. The White House accused Chinese entities of running industrial-scale distillation campaigns in late April. Datasets containing Claude Opus 4.6 reasoning outputs are circulating on HuggingFace with unclear provenance, highlighting the scale of the data harvesting operation.