Token brokers advertise AI inference at 30% to 80% discounts through relay proxies
- A broker told the author it could support $100,000 per day in spend, supplied an API endpoint rather than provider credentials, and said it would bill after a usage milestone; the author infers that the service forwards requests through a pool of keys.
- The AI Credits marketplace lists OpenAI, Anthropic, Gemini, Azure, ElevenLabs, and MiniMax credits at 30% to 80% discounts, while its seller form asks sellers to choose a provider, credit type, value, discount, and delivery method.
- The author submitted pending listings for $200,000 in OpenAI credits and $10,000 in Anthropic credits on AI Credits, showing that the marketplace accepts large proposed resale listings before approval.
- CheapCredits advertises a flat 40% discount on GPT-5 input and output pricing and says its routing savings come from bulk pricing; the author says that discount is unlikely without top-tier provider volume and suspects another supply source.
- CheapCredits publishes a data processing agreement that claims GDPR Article 28 compliance and names OpenAI and Anthropic as sub-processors, despite routing customer requests through its service.
Hacker News opinions
I think Yunwu.ai was a way to get ultra-cheap tokens while someone might inspect your traffic. They dropped US models, and now I wonder whether startup credits are being turned into cash.
I've seen thousands of similar sites using the newapi core in Chinese forum linux.do. They were never passing through real US models in some cases: testing suggested their supposed Anthropic responses were mediocre Kimi-based distillations.
I suspect supply can come from stolen cards, demo accounts, free trials, leaked company credentials, and unlimited ChatGPT-style relays.
Startup grants are another obvious source. Providers give startups tens or hundreds of thousands of dollars in credits, and a startup can trade unused credits for cash.
Unused-credit trading sounds more plausible than fraud at modest discounts, even if it breaks the provider agreement. But discounts up to 98% point to stolen API keys, stolen cards, mass trial-account creation, or a substituted model instead of the requested API.
Claude Max 20x costs $200 yet can amount to thousands of dollars of API-equivalent use. Resellers can also profit from kickbacks tied to token history collected for distillation.
Cheap tokens are a very effective way to collect data for distillation. I would be surprised if that is not part of the business model.
A reseller proxy can terminate TLS and alter LLM tool calls before the client executes them. That can enable secret exfiltration or client-machine control, depending on the harness permissions.
I do not see evidence that all of this is fraud. A plain resale market is a simpler explanation for at least some listings.
I have spoken with companies seeing this, and some cases are credit-card fraud, chargebacks, and account takeovers. Tokens have become a pseudo-currency, so both benign resale and real abuse exist.
I understand why someone with free credits wants to sell them. I still want to know who buys them and why they accept the proxy and provenance risk.