How China’s Shadow AI API Market Works - Vincent Schmalbach

Skip to content

Currently Available: Need a skilled Software Developer for your next project?

Hire Me Today

Categories

Economics LLM

How China’s Shadow AI API Market Works

May 18, 2026<br>by Vincent Schmalbach

China's shadow market offers access to Claude, Gemini, GPT, and other frontier models. Pay a local seller, get an API endpoint, connect it to a coding tool, and use models that are hard or impossible to reach directly from mainland China.<br>I'm writing this while working on vroni.com, my AI coding tool for turning tasks into pull requests.

These API relay platforms are being advertised on Taobao and Xianyu. Sellers promise no-VPN access, low latency, large context windows, coding-tool compatibility, and official-looking Claude, Gemini, and ChatGPT access. Some listings claim "1:1 official models."

Claude and Gemini are not normally available in mainland China. Those shadow API sellers offer a workaround: instead of calling Anthropic, Google, or OpenAI directly, the developer calls the seller's server.

That server is the middleman. It receives the prompt, sends it somewhere else, gets an answer back, and returns that answer to the developer's tool.

The product is a working API path to models the buyer cannot easily call directly.

What you buy

The seller usually gives the buyer three things:

an API base URL

an API key

a list of model names the endpoint claims to support

To the buyer, it looks similar to using an official API. Put the base URL and key into your tool (e.g. a coding harness), choose a model name, and send requests.

The important difference is that the endpoint is not run by the model company. It is run by the seller. The buyer is not sending prompts directly to Anthropic, Google, or OpenAI. The buyer is sending prompts to a third party that decides what to do next.

The possible upstream routes can be an official API account, a cloud account, a consumer subscription, a pool of accounts, or a cheaper substitute model returned under a more expensive model name.

As a user you cannot verify which route the seller actually used for a request.

Where the request goes

The flow is straightforward:

Developer tool<br>-><br>Seller's API endpoint<br>-><br>Upstream account or model chosen by the seller<br>-><br>Answer returned to the user<br>The seller's endpoint is doing the routing. That means the seller controls the upstream account, the model choice, the logs, rate limits, and any fallback behavior.

A March 2026 audit of shadow APIs found weak transparency around provider identity, upstream models, and infrastructure. The authors identified 17 shadow API providers that had already appeared in research and open-source workflows.

So the model name you are seeing as a user might very well be fake. A dropdown that says Claude or Gemini does not prove that the request actually went to the official Claude or Gemini API.

Pricing below official API prices

Some sellers advertise access below official API prices.

Small discounts can have ordinary explanations: unused quota, promotional credits, volume pricing, or a cheaper payment path.

Very large discounts need a different supply source. ChinaTalk's investigation into cheap Claude tokens in China describes transfer stations, account merchants, SMS verification services, card merchants, proxy networks, subscription pooling, and downstream resellers. It says some Claude access is sold at roughly 10% of the official price.

Several mechanisms can reduce the seller's cost:

harvest trial credits

abuse educational or corporate discounts

turn subscription access into shared API-style access

use stolen credit cards

route traffic to cheaper models

Coding agents make subscription resale especially attractive. A normal subscription is priced for one user. If a seller turns that subscription into a shared endpoint for many users, the apparent per-user cost drops until the account is limited or banned.

Anthropic's February 2026 distillation report shows the same kind of account infrastructure at larger scale. Anthropic said DeepSeek, Moonshot, and MiniMax generated more than 16 million Claude exchanges through about 24,000 fraudulent accounts. It also described proxy networks where banned accounts are replaced and traffic is spread across many nodes.

The model might be fake

A shadow API seller can advertise one model and serve another.

If the buyer asks for Claude Opus or Gemini Pro, the seller can send the prompt to a cheaper model and still return the response under the expensive model name.

The shadow API audit found performance gaps, inconsistent safety behavior, and failed fingerprint checks when comparing shadow APIs with official APIs. In one reported case, an endpoint sold as Gemini-2.5 performed far below the official API on a medical benchmark.

This matters for normal users, but it also matters for research. If a paper or benchmark uses a shadow API endpoint while assuming it is...