Shadow APIs: Bypassing Restrictions to Access Claude and Gemini

The Rise of Grey Market AI Relay Platforms in China

In recent years, a thriving grey market has emerged in China, offering developers access to top-tier overseas AI models such as Anthropic's Claude and Google's Gemini. These models are officially unavailable in the country due to regulatory restrictions, yet demand remains high among local users seeking advanced capabilities for tasks like coding, debugging, and image generation.

Relay platforms have become a go-to solution for developers who want to bypass these restrictions. These services route access to overseas AI models through proxy servers hosted outside mainland China, enabling users to engage with powerful tools that are otherwise inaccessible.

On Chinese online marketplaces such as Taobao and Xianyu, relay providers are advertising native Claude Opus access, unlimited Claude Code subscriptions, and 1:1 official models without any capability reduction. Most sellers promote support for one-million-token context windows, domestic network access without the need for a VPN, and compatibility with tools like Cursor, VSCode, and OpenClaw.

One high-volume seller on Xianyu, who has fulfilled more than 2,200 orders, advertised "low-latency, no-VPN" access to the full Claude 3.5 suite. Online listings advertising access to Claude, ChatGPT, and Gemini remain common across the two marketplaces, alongside second-hand electronics and gaming hardware.

The demand for these underground services is driven by a persistent performance gap between Western AI leaders and domestic alternatives. Even as Chinese models become cheaper and more widely available, many developers still prefer the accuracy and reliability of US-based AI systems.

"Claude's outputs are usually very accurate," said a Hangzhou-based programmer surnamed Song, who regularly relies on the restricted US model for complex engineering workflows. "Even when my instructions are not completely clear, it can still execute tasks well. I rarely need to fix bugs afterwards. Chinese models still hallucinate more often and sometimes generate code for things I never asked for."

Global Expansion of AI Relay Infrastructure

The underlying relay infrastructure is also becoming mainstream globally. In early May, Chinese crypto entrepreneur Justin Sun announced on X that he was launching his own AI relay platform. Shortly after, WLFI, a US-based cryptocurrency company backed by the Trump family, unveiled a similar service called WorldRouter, promising users access to more than 300 global AI models through a unified API interface.

However, resellers are facing increasingly tighter restrictions imposed by model developers. Several API resellers said the business had become harder to operate after Anthropic tightened account enforcement last month.

US AI companies have progressively tightened access for users in mainland China over the past year. Anthropic, best known for its Claude models, has gradually expanded controls covering mainland China, Hong Kong, and Macau across Claude web access, APIs, and developer tools including Claude Code. Access already required overseas phone numbers, foreign payment cards, and non-Chinese billing addresses.

Anthropic further restricted access for entities majority-owned by organizations based in unsupported regions, including China, last September. The company also introduced identity verification checks for some users, requiring government-issued identification and real-time selfie verification through identity platform Persona in mid-April.

"Relay services were still relatively easy to operate before mid-April," said a 27-year-old reseller surnamed Zhao, who sells Claude access through second-hand marketplace Xianyu. "After that, account bans became much more serious, especially with the new KYC (know your customer) checks. Once an account is banned, the subscription fees stored in it are effectively lost."

Cost and Quality Concerns in the Relay Market

Many relay platforms advertise API prices below official rates. One million tokens of GPT-5-level API usage can cost substantially less than official rates. KoalaAPI.com advertised GPT-5.4 access at around US$1.50 per million tokens, compared with the ChatGPT official rate of US$2.50 per million input tokens. Another relay platform, Xinglian 4SAPICOM, advertised prices as low as US$0.36 per million tokens under what it described as smart routing, according to the company's official website.

A Beijing-based provider surnamed Zhang, who has sold over 1,000 orders for Claude relay services with an 83 per cent positive rating, said users may not always know which model was processing their requests. Some relay operators advertise access to overseas frontier models while dynamically substituting requests with lower-cost Chinese systems such as MiniMax or Qwen.

"Most Chinese users do not have enough direct experience with frontier overseas models to reliably tell what they are actually using," Zhang said. "The market has become heavily diluted, and lower-quality operators are beginning to drive out the more reliable ones."

Another seller advertising Claude access said he built his own relay cluster after becoming frustrated with unstable unofficial interfaces that frequently triggered rate limits or corrupted long-context conversations.

Security Concerns and Regulatory Actions

In April, the White House warned that Chinese entities were conducting industrial-scale distillation attacks against advanced US AI systems through large networks of proxy accounts designed to evade detection. Anthropic also disclosed attempts by China-linked actors to access its models through coordinated intermediary infrastructure in February 2026.