When Western AI labs draw digital borders around their models, they do not create a vacuum. Instead, they trigger a hyper-efficient, highly profitable, and slightly chaotic shadow economy.
If you are a developer or researcher living in China, accessing frontier models like OpenAI’s GPT-5 or Anthropic’s Claude 4 is a constant battle. Officially, these services are geoblocked. Unofficially, a quick search on e-commerce giants like Taobao or second-hand platforms like Xianyu reveals thousands of listings offering premium access for pocket change.
I wanted to know how this underground ecosystem actually operates. To find out, I pulled out my wallet, bought access from multiple different vendors, and put their services to the test. Here is the investigative anatomy of China's "Transfer Station" economy, along with the results of my hands-on testing.
The Three Tiers of the Grey Market
The underground market for Western AI is not a monolith. It is a highly segmented hierarchy catering to different budgets and technical needs.
Tier 1: Shared Web Accounts (The Budget Option): Perfect for students and casual users. On Taobao, you can buy login access to a single Claude Pro or ChatGPT Plus account that is being shared by 10 to 50 people. If too many users prompt the AI at once, you get hit with rate limit errors, and you can see everyone else's chat histories in your sidebar.
Tier 2: The "Transfer Station" (中转站, or Zhōngzhuǎnzhàn): For developers and startups, shared web accounts are useless. They need stable API keys. This need created Transfer Stations (API proxy services that act as intermediaries). Instead of pointing their code to the official endpoints (which require foreign credit cards and VPNs), developers change their base URL to a custom proxy link. The transfer station accepts local payments like WeChat Pay or Alipay and routes the requests through overseas servers.
Tier 3: Fake Identities & KYC-Bypassed Accounts: When Anthropic began partnering with third-party verification platforms like Persona (requiring select users to upload physical government photo-IDs), the market adjusted instantly. Sellers now offer fully registered accounts backed by fake or stolen international IDs, turning verification bypasses into a premium commodity.
Real-World Testing: What I BoughtWhen I logged into the purchasing dashboards of these services, I was met with highly structured menus offering various tiers of access. They call these shared accounts "fleets" or "shared carriages" (车队).
The packages are highly segmented. You can buy cheap weekly passes for about ¥45 (around $6 USD) or monthly packages for ¥180 (around $25 USD).
If you look closely at the cheap tiers, they explicitly state "不可用Fable5" (Fable 5 not available). To get access to the absolute highest-tier frontier models, you have to shell out premium money, all the way up to ¥1500 (around $200 USD) for a dedicated, high-limit monthly fleet account.
The Scams: Kimi in a Trench CoatMy first stop was testing the ultra-cheap sellers. I wanted to see if their custom web portals were actually running the real Claude models.
It did not take long to spot the fakes.
When I started prompting the cheaper web interfaces, the AI's writing style felt off. It was too fast, unusually polite in a very specific corporate-Chinese way, and failed classic logical reasoning tests that Claude Sonnet usually passes with ease.
By engineering a few specific system-prompt bypasses, the truth came out. These cheap vendors were not using Claude at all. They were running local, inexpensive Chinese open-source models (like Moonshot's Kimi or Alibaba's Qwen) through custom transformers. They wrap these models in a Claude-themed user interface, tricking casual users into thinking they are getting premium Western AI for pennies.
If you pay bottom-dollar on these platforms, you are almost certainly getting a rebranded local model.
The Real Deal: Unlocking Claude Fable 5However, the higher-tier vendors were entirely legitimate. When I bought access to their premium API proxies, I routed my development environment through their custom transfer station URLs to monitor the actual token usage and response logs.
The logs confirmed that the API was actively routing and billing tokens for claude-fable-5. The response latency (averaging between 8 and 14 seconds for first-word delivery) and the sheer depth of the reasoning proved this was the genuine, unadulterated flagship model.
Taking it to the Terminal: Claude Code
The real test for any developer is getting local terminal tools to work. I configured my system to run Anthropic's command-line tool, Claude Code, using the proxy credentials and the guide I purchased.
It worked flawlessly.
I was able to run Claude Code locally, index my directories, and let the agent write and refactor code directly inside my terminal, completely bypassing the geographic blocks (used a vpn to test where claude is blocked)
The Dark Math: How is it 90% Cheaper?
The most mind-boggling part of this economy is the pricing. Some transfer stations sell Claude tokens at rates significantly below official Anthropic prices. How can a reseller turn a profit while selling a premium product so cheaply? They use a mix of clever engineering, arbitrage, and outright deception:
The Silent Bait-and-Switch: If you ask a proxy a simple question like "Who was the first president of the United States?", it does not waste money sending that query to Claude 4 Sonnet. The proxy silently routes simple questions to a dirt-cheap local model. You get your answer instantly, and the reseller pays next to nothing.
Prompt Harvesting: This is the real danger. Some operators log every prompt, codebase, and response sent through their proxies. They pack these logs into massive data dumps and sell them to local AI firms hungry for conversational training data.
Stolen and Trial Credits: Many operators buy pooled accounts set up using stolen credit cards or heavily subsidized developer credits, dumping them on the market before the accounts get flagged and banned.
The Cat-and-Mouse Timeline
This grey market did not emerge overnight. It has evolved step-by-step alongside the AI boom:
Late 2022: The Sandbox Era
The Catalyst: OpenAI releases ChatGPT to the world.
The Workaround: The Great Firewall blocks are easily bypassed using basic commercial VPNs. Chinese users sign up using temporary online SMS-receiving platforms to catch verification codes.
2024 to 2025: The Great Clamping
The Catalyst: Anthropic and OpenAI start actively blocking known VPN IP ranges and rejecting virtual credit cards (like Dupay).
The Workaround: Taobao sellers pivot. Instead of selling individual DIY accounts, they begin selling shared browser logins and pre-activated "fleet" accounts hosted on clean overseas servers.
April 2026: KYC Escalation
The Catalyst: Anthropic implements strict photo-ID verification using third-party verification systems to weed out bot accounts.
The Workaround: The black market adapts within days. Sellers begin sourcing fake or stolen foreign IDs to bypass verification, causing a massive price premium for guaranteed "clean" accounts.
Mid-2026: The Proxy Standard
The Catalyst: Web accounts become too unstable to maintain for serious developers.
The Workaround: The birth of the professional "Transfer Station" ecosystem. Instead of browser logins, developers use custom API proxy URLs. It is incredibly convenient, but users are quietly trading away their proprietary code and private data to operators who harvest and resell it.
The Takeaway: Borders Do Not Stop Demand
The cat-and-mouse game between Western AI providers and the Chinese grey market shows no signs of slowing down. For engineers in China, access to elite models like Claude for coding tasks is essential, and they are willing to leap over security fences, risk bans, and ignore data leaks to get it.
Just remember: when you use a shadow proxy, you are trading your privacy and data security for a cheaper token. If you are working on proprietary code, the discount is never worth the risk.