A gray market resells stolen API access specifically to distill frontier models

AI-generated writing.

There’s a whole supply chain built around reselling stolen or bulk-fraudulent access to frontier model APIs in mainland China. Card and account merchants sell virtual credit cards and bulk-registered accounts, “account pools” aggregate hundreds of those accounts behind a single API, and consumer-facing “relays” wrap that pool in a billed, Chinese-language product — undercutting official pricing by as much as 97.8%. One sampled package bought $3,333 of official Anthropic credit for 425 RMB, about $0.13 of real usage per $1 spent.

One of the named buyer segments for this cheap access is model distillation: companies collecting large volumes of outputs from Claude and other frontier models specifically to train their own domestic models. A forum source described it as a multi-billion RMB industry with some operators earning hundreds of thousands of RMB a day. It’s a concrete, industrial-scale version of the distillation-as-theft scenario — rather than one lab quietly querying another’s API, there’s now dedicated infrastructure whose purpose is laundering identity and payment fraud into cheap bulk access for exactly that purpose.

The market is mature enough to have price-comparison sites, affiliate programs, and even a daily lottery giving away $100 API keys — with a “provably fair” cryptographic draw borrowed straight from crypto-gambling sites. That kind of normalization, more than the technical mechanics, is the part worth remembering: this isn’t a handful of opportunists, it’s an established ecosystem with its own tooling (mostly built on two open-source gateway projects, one-api and new-api) and its own reputational infrastructure for trust between fraudulent actors.