The discounted pricing is sustained by stolen credentials, model substitution, and harvesting users’ prompts and outputs for resale as AI training data, enabling further monetization.

A Chinese gray-market ecosystem resells access to Anthropic's Claude AI at roughly 10% of official prices through proxy networks called transfer stations, openly operating on platforms like GitHub, Taobao, and Telegram.

These findings echo warnings from the White House and Anthropic about industrial-scale distillation campaigns targeting U.S. frontier models using tens of thousands of proxy accounts and fraudulent activity.

Overall, the piece underscores the tension between affordable access to powerful AI tools and the ethical, legal, and data ownership implications of unregulated discount channels.

Oxford China Policy Lab researcher Zilan Qian provides concrete evidence of how transfer stations operate and monetize access, fueling concerns about data security and AI model integrity.

These transfer stations resell API tokens to enable data extraction and model distillation, effectively allowing mass queries to Claude.

The practice is likened to buying a luxury item at a deep discount only to discover the seller uses parts to build a rival product, highlighting risks and the lack of guarantees about data handling and model access.

Model distillation is the technical strategy involved, where outputs from large models are used to train smaller, more efficient systems, raising concerns about data ownership and unauthorized use of commercial AI models.

Users purchasing Claude through proxies may have their data repurposed to improve or train the buyers’ competing models, raising ethical and legal questions around privacy and IP.

Anthropic has limited ability to control the distribution of its API tokens once resold, highlighting governance and enforcement challenges in the AI marketplace.

Anthropic previously disclosed nearly 24,000 fraudulent accounts tied to Chinese labs aiming to copy or distill frontier models, signaling broader fraudulent activity.