How Model Transfer Stations Exploit Token Economics – The “One Fish, Three Ways” Model

Claude Opus 4.8 officially costs $5 per million input tokens and $25 per million output tokens, roughly ¥34 and ¥170 respectively; yet many Chinese marketplaces sell the same output for as little as ¥5 per million tokens, a discount of over 90%.

A “transfer station” is an API proxy that sits between users and providers such as Anthropic or OpenAI, forwarding requests while masquerading as a legitimate overseas account, allowing payment via WeChat or Alipay without a foreign credit card.

The service exists because Anthropic and OpenAI increasingly restrict access: Chinese users cannot register without an overseas phone number or credit card, corporate ownership rules were tightened in September, and stricter identity verification began in April, prompting a market for work‑arounds.

Paradoxically, the tighter the restrictions, the larger the middle‑man business; Singapore’s per‑capita Claude usage tops global rankings, and Chinese developers jokingly claim they become “temporary Singaporeans” to obtain access.

Why can transfer stations undercut official pricing? The author frames the economics as a three‑course meal:

First course – arbitrage. Operators acquire Anthropic’s $5 registration credits in bulk, resell unused quota, exploit enterprise or education discounts, or split a $200 “Max” subscription among multiple users (APImaxxing). Some even use stolen credit cards to create zero‑cost accounts, driving prices below ¥1 per dollar of token.

Second course – model mismatch. The proxy can silently route a request labeled as Opus 4.8 to cheaper models such as Sonnet, Haiku, or domestic alternatives like GLM, Qwen, or DeepSeek, saving four to eight percent of costs. CISPA (German Information Security Center) audited 17 API proxies and found widespread model substitution; a claimed Gemini‑3 proxy achieved only 37% of the official API’s 83.82% score on a medical benchmark.

Some operators employ dynamic routing: high‑value users or short prompts may reach the genuine Opus model, while peak‑time, long‑context, or low‑price requests are diverted to cheaper models, making the profit model highly adaptable.

Third course – data resale. Every request’s full prompt, response, tool calls, and iteration history are stored on the operator’s servers. For coding agents, this data contains real engineering decisions, long reasoning chains, repository context, and verified answers—high‑value material for supervised fine‑tuning or distilling Claude’s reasoning into smaller models. Unattributed Claude Opus datasets already circulate on HuggingFace, likely sourced from such transfer stations.

The ecosystem’s supply chain is highly compartmentalized. Upstream, “account merchants” bulk‑register or purchase accounts, provide overseas phone numbers, and reverse‑engineer Anthropic’s client code to bypass authentication. KYC is cracked using AI‑generated documents, deepfakes, or by hiring individuals in Africa, Latin America, or Southeast Asia for on‑site verification; black‑market iris scans from Worldcoin sell for under $30 each. The middle layer is the transfer station itself, while downstream are developers, application owners, and secondary resellers who wholesale tokens on platforms like Taobao.

Because no single actor controls the entire chain, the system is resilient: if authorities shut down one operator, the upstream account pool and downstream customers enable a new proxy to emerge within hours.

Legally, transfer stations operate in a grey zone in China. Providing AI services without a备案 (record) or security assessment violates regulations, but enforcement is lax until the business scales. Recent alerts from the Ministry of State Security and the Cyberspace Administration warn of unqualified platforms, weak security, privacy leaks, and data resale risks.

Thus, the ultra‑low price is sustained not by altruism but by extracting value through arbitrage, model substitution, and especially by monetizing user conversation data. Sustainable operators focus on integration capabilities—unified interfaces, logging, regional routing—exemplified by platforms like OpenRouter.