An investigation published on July 26, 2026 describes the „relay market” — a gray market where resellers, mostly in China, sell cheap access to LLM tokens. They pool API keys obtained from free trials, stolen cards and chargebacks. The analysis was done by Matt Lenhard of Vectoral, based on a thread on the Chinese forum v2ex, and amplified by Simon Willison.
Key takeaways
- The market runs mainly in China and relies on proxies that pool API keys from many providers.
- Resellers obtain tokens from free trials, unprotected support bots, stolen cards and chargebacks.
- The most common tools are the open-source one-api and its more actively developed fork new-api.
- Buyers want discounts, ways around geographic blocks, and data for model distillation.
- The investigation's main source is a v2ex forum thread, reported by Matt Lenhard of Vectoral.
What the token „relay market” is
The relay market is a network of intermediaries who resell access to language models cheaper than the official providers.
The mechanism is simple: a reseller runs a proxy server, loads it with a pool of other people's API keys and exposes their own endpoint.
The client pays the reseller, and the traffic goes to the real providers at the expense of the owners of stolen or abused keys. To the buyer it looks like an ordinary API, only cheaper.
This inverts the model we know from legitimate gateways and routers, such as the aggregator OpenRouter.
There, the intermediary has contracts and settles with providers. In the relay market the whole economy rests on someone else paying the bill — or nobody paying it until the provider detects the abuse.
The market is concentrated in China, where access to some Western APIs is restricted and demand for cheap compute is high. That explains why the main source of the analysis was a Chinese-language thread on the v2ex forum.
Where the cheap tokens come from
According to the investigation, resellers acquire keys and quotas in several ways. The mildest is mass account creation to reuse free trials many times over.
The harsher methods are proxying through unprotected support bots, paying with stolen cards, and chargebacks?Chargeback: A card payment reversed by the buyer's bank, after the service has already been delivered. — reversing a payment after the tokens have already been spent. In practice the provider bears the inference cost and never gets paid.
Detecting such abuse is hard, because proxy traffic looks like ordinary requests from many accounts. The provider sees scattered, individual calls rather than one suspicious stream.
Before it blocks a stolen key, the reseller has already monetized it and moved on to the next one in the pool. It is a cat-and-mouse game in which the defender always reacts late.
The tools: one-api and new-api
The technical backbone of the market is two open API-proxy projects. The first is one-api. The second, new-api, is its fork, described as more actively developed.
Both let you manage a pool of keys, balance traffic across them and expose a single endpoint compatible with the format of popular providers. In themselves they are neutral tools — the problem is what you feed them with.
Who buys and why
Buyers have three motivations. The first is simply price — access below list price. The second is getting around geographic blocks where the official API is unavailable.
The third, the most technically interesting, is collecting data for model distillation — training a smaller model on a larger one's outputs. A cheap, anonymous pipe to a frontier LLM is a convenient tool for this, because it lets you generate large sets of question-answer pairs without paying full price and without leaving a trace on your own account.
Why it matters
The relay market shows that billing for LLM access is still a weak link. An API key is real money today, yet many providers do not offer simple, hard spending caps per key.
The effect is twofold: companies lose money on unpaid inference, and at the same time their models become raw material for training competitors via distillation. That shifts the problem from „someone is stealing compute” to „someone is siphoning off our model advantage”.
There is also a systemic dimension. As long as free trials, poorly protected endpoints and easy chargebacks remain the norm, the gray zone will pay off. Willison puts it bluntly.
LLM vendors really need to get better at offering strict caps for their API keys.
— Simon Willison, blogger and Django co-creator.
What's next
- Pressure on LLM providers to introduce hard spending caps per API key — Willison's main takeaway from the investigation.
- The risk that cheaply acquired outputs will be used to distill competing models — an angle that goes beyond mere compute theft.
- The discussion continues on the v2ex forum and in a Hacker News thread, where specific tools and abuse methods are analyzed.
Sources
- Vectoral — An Inside Look at the Relay Market Powering Token Resellers and Fraud
- Simon Willison's Weblog — An Inside Look at the Relay Market Powering Token Resellers and Fraud
