Grey-market proxies resell Claude API access at discount

A report by Oxford China Policy Lab researcher Zilan Qian, as covered by Tom's Hardware, finds a grey-market ecosystem of API proxy services in China reselling access to Anthropic's Claude models at prices reported as low as 10% of official rates. The report says these proxy operations, known locally as "transfer stations," are visible on platforms including GitHub, Taobao, and Telegram and implement a modular supply chain where participants specialise in one or two links. Qian's reporting describes tactics including bulk account registration by farming free API credits, exploiting corporate discounts, subdividing $200 Max subscriptions across many users, and introducing accounts purchased with stolen credit cards. The article notes the proxies sustain margins by quietly substituting different models and logging both user prompts and model outputs for resale as training data.

Editorial analysis - technical context: The practices described create two technical risks commonly discussed in security and ML operations. First, model-substitution and proxying break the trust boundary between an API client and an advertised model, producing unpredictability in latency, correctness, and the provenance of outputs. Second, systematic collection of prompts and outputs by intermediaries raises data-leakage and dataset-contamination concerns, since those interactions can be ingested into downstream training pipelines without user consent.

Tom's Hardware links Qian's findings to recent public warnings: the White House alleged in late April that Chinese entities have run "industrial-scale" distillation campaigns against frontier models, and Tom's Hardware reports that Anthropic previously disclosed account misuse tied to groups including DeepSeek, Moonshot AI, and MiniMax. For practitioners, these reports underscore operational exposure when routing traffic through third-party proxies and the legal and compliance risks of unintentionally contributing proprietary prompts or outputs to external training sets.

indicators include takedowns or policy enforcement by hosting platforms where "transfer stations" advertise, vendor disclosures of unusual API call patterns or account aggregations, and changes to API pricing, rate limits, or credential-issuance controls intended to curb bulk account farming. Observers should also track any public incident reports from enterprise users whose prompts may have been captured and reused.