An Inside Look at the Token Reseller Market
An Inside Look at the Token Reseller Market
By Matt Lenhard | June 28, 2026
My fascination with token fraud began during my tenure as a software engineer working on an AI gateway. Initially, the issue seemed simple: free-credit abuse, where malicious actors created thousands of accounts to harvest trial credits. However, as I consulted with colleagues, I discovered that companies were hemorrhaging millions of dollars daily.
The attackers were not just opportunistic; they were relentless. A systemic form of fraud had embedded itself into the very fabric of the token economy. While investigating the source of this traffic, I discovered a Chinese forum where operators openly detailed the mechanics of relays (also known as "transfer stations").
The Economics of the Relay
A relay acts as a proxy service that routes traffic to U.S.-based AI providers. The price discrepancy is staggering. For instance, one operator offered a package providing the equivalent of $3,333 in official Anthropic credits for a mere 425 RMB.
To put this in perspective using :
This represents a 100% price point 97.8% discount compared to official rates.
Top Relay Providers by Discount
The following table highlights the most aggressive price-cutters we tracked:
| Rank | Provider | Median Discount from List Price |
|---|---|---|
| 01 | Now Coding | 97.8% |
| 02 | I Code Easy | 97.1% |
| 03 | Claude ZZ | 96.6% |
| 04 | Doro | 96.4% |
| 05 | UoCode | 96.3% |
| 06 | ZeroCode | 94.9% |
| 07 | AiYa | 94.9% |
| 08 | HongMaCC | 94.2% |
| 09 | Right Code | 94.1% |
| 10 | BUZZ | 94.1% |
| (Total providers tracked: 49) |
Market Architecture: The Four-Layer Stack
The ecosystem is structured as a deep supply chain, moving from raw account sourcing to the end developer.
01. Upstream (卡商 / 号商)
These are the Card and Account Merchants. They specialize in providing bulk-registered accounts and virtual credit cards designed to bypass U.S. and European billing verification systems.
02. Midstream (账号池)
Account Pools act as the aggregators. They take hundreds of upstream accounts and manage the "messy" side of the operation:
- Managing authentication tokens.
- Handling rate limits.
- Implementing failover logic when accounts are banned.
- Exposing a unified API for downstream buyers.
Interestingly, these pools don't just use official APIs. They often use reverse-engineered access to consumer apps (like Kiro or antigravity). If a model is exposed via any interface, the pool will find a way to scrape or proxy it.
03. Downstream (中转站)
The Relays are the customer-facing storefronts. They wrap the pool's API in a polished, Chinese-language product. Their primary focus is commerce: billing, invoicing, and managing customer support via WeChat.
04. End Users
The final consumers include:
- Individual developers and startups seeking cheap inference.
- Mid-sized SaaS companies.
- Commercial entities performing model distillation.
The Technical Engine
Most relays rely on two primary open-source projects: one-api or new-api.
The operational flow looks like this:
- The operator deploys the panel.
- They configure channels (渠道) linked to the account pools.
- The panel provides a single OpenAI-compatible endpoint.
- The system handles the request logic:
# Conceptual logic of a relay request
def handle_request(user_request):
api_key = pool.get_available_key() # Rotate keys to avoid rate limits
response = forward_to_provider(api_key, user_request)
# Calculate cost based on a multiplier (倍率)
cost = response.tokens * multiplier
user_account.deduct_quota(cost)
return response
While one-api is the more common foundation, new-api is a fork that adds advanced commercial features like self-service recharging and support for multi-modal (image/video/audio) models.
The Fraud Playbook
A relay becomes fraudulent when it resells access using stolen or abused keys. Here is the "checklist" of techniques used by abusers:
- Credit Harvesting: Automating mass account creation to claim free trial credits.
- Financial Fraud: Using stolen credit cards or initiating chargebacks after the tokens are consumed.
- Limit Capping: Funding accounts with low-limit prepaid cards.
- App Proxying: Reverse-engineering consumer chatbots to route traffic through them.
- Spend Burning: Launching massive concurrent request floods specifically to deplete a provider's budget.
The Ultimate Goal: Model Distillation
Beyond just saving money, a significant driver for this market is model distillation. This involves using high-end frontier models (like Claude) to generate synthetic data to train smaller, domestic models.
"Distillation uses Claude/CodeX models to train domestic models... companies with strong programming capabilities are all distilling Claude." — Translated from V2EX forum
The scale of this industry is immense, with some distillers reportedly making millions. Even the marketing for these services is gamified, as seen in the daily lotteries for API keys.
