Turns Out There's a Huge Black Market for Claude AI Tokens in China and It's Selling at Up to 93% Cheaper

A sprawling grey market for Claude AI tokens has emerged in China, where unofficial resellers are reportedly offering access to Anthropic's flagship models at discounts of up to 93% below official API prices.

The claims, raised in a widely discussed Hacker News forum thread and supported by a detailed analysis from researcher Zilan Qian, suggest the underground ecosystem extends far beyond simple account sharing, relying instead on a sophisticated network of proxy operators, account suppliers, and, in some cases, alleged data harvesting.

The discussion followed Anthropic's recent claims that Chinese entities had used large networks of proxy accounts to extract data from its models. The company has tightened access controls over the past year by expanding identity verification requirements, adding biometric checks for some users, and restricting access from unsupported regions, including mainland China.

Yet, researchers and developers argue these measures have also fuelled a parallel economy designed specifically to bypass those restrictions.

Claude AI Tokens Are Fueling a Hidden Resale Economy

According to the forum discussion, the unusually low prices stem from multiple overlapping business models rather than a single source of discounted access.

One of the most common methods involves pooling subscriptions, such as Claude Max accounts, and redistributing unused capacity among dozens of customers. Others reportedly rely on bulk registration campaigns, promotional credits, educational discounts, or corporate pricing that can be repackaged and sold at scale.

The more controversial allegations concern payment fraud and user data. Participants in the discussion claimed some operators use accounts created with stolen payment details, although no evidence was presented to establish how widespread that practice may be.

The largest financial incentive may not be token sales at all.

Zilan Qian's research argues that many proxy operators treat API access as a customer acquisition tool rather than their primary source of revenue. Instead, every prompt, response, and coding session that passes through their servers could become valuable training data.

Chinese developer communities have claimed that at least some operators collect logs that can later be used for AI model training or sold to third parties, though the research notes that systematic harvesting and commercial sales remain unconfirmed.

That distinction matters. While discounted access can attract users, the long-term commercial value may lie in the conversations themselves.

A Growing Challenge for Anthropic

The pricing alone illustrates why the market has attracted attention.

One Hacker News contributor claimed a reseller was offering access to Claude Opus at roughly 93% below Anthropic's official API pricing. Other participants disputed parts of the explanation, arguing that cheaper open-weight models such as DeepSeek and GLM are primarily inexpensive because they are more computationally efficient rather than because they are competing directly with Claude resellers.

Even so, several developers agreed that unofficial proxy services have become difficult for legitimate providers to ignore.

One contributor argued that users in mainland China face several practical barriers to registering directly with Anthropic. In addition to requiring a VPN, the service generally cannot be paid for with Chinese bank cards, and identity verification requirements may prevent many users from keeping their accounts active. Those hurdles have created demand for intermediaries that accept domestic payment methods such as WeChat Pay and Alipay while routing requests through overseas infrastructure.

Not everyone accepted the central claims without challenge.

Several commenters questioned whether many resellers are actually delivering genuine Anthropic models, suggesting that some could instead be substituting Chinese models while marketing them as Claude. Others pointed to research cited by Qian showing that certain API proxies substantially underperformed official services, potentially indicating undisclosed model switching. Users often have little way to verify which model processed their request.

The Problem With AI Subscription Plans

The discussion also exposed the loophole behind AI economics.

Some contributors argued that heavily subsidised subscription plans effectively invite arbitrage, while others described the resale market as straightforward fraud that violates Anthropic's terms of service.

Anthropic has not publicly confirmed the scale of the reported resale ecosystem or responded directly to many of the specific allegations discussed in the forum. Likewise, claims surrounding payment fraud, systematic log harvesting, and the resale of reasoning traces remain largely based on industry discussions and researcher interviews rather than independently verified evidence.

For developers, however, the debate extends beyond cheap API calls. It raises difficult questions about data privacy, platform security, and whether increasingly strict access controls are unintentionally creating an even larger underground market that profits from the very restrictions they were designed to enforce.