A large, opaque resale market for frontier LLM APIs has emerged to serve users blocked from OpenAI, Anthropic and Google endpoints (somewhat like a VPN). Two independent papers in the past two months provide the first systematic evidence that a substantial fraction of these intermediaries silently substitute models, truncate context, or misreport billing.
Apart from inflated price, the shadow market maybe more chaotic than ideal: for example, shadow-API providers would steal tokens by using fake educational accounts. The claim to provide expensive good models while quietly substituting for cheaper ones.
And there’s even more. When Anthropic started real-ID-verification, shadow API providers pivoted to poor countries to get headshots and fake ID for verification—a few dollars paid to the locals—and later sell eg Claude API at 7x price.
For users, using a crap model is fine but sometimes the shadow-API AI-agents are intentionally seeding malicious software on their computers when granted system access. Not to mention the lost of data privacy.
Given the market’s size and profitability (mainly software engineering), why am I not surprised?
Some interesting reference to look into:
[1] APPSO. “AI scalpers — how deep is this lucrative business?” 8 May 2026. https://mp.weixin.qq.com/s/SmmAVhILzwASP8tpuNGEKw
[2] Zhang, Y., Jiang, Y., Chen, Z., Backes, M., Shen, X., Zhang, Y. Real Money, Fake Models: Deceptive Model Claims in Shadow APIs. CISPA Helmholtz Center, arXiv:2603.01919v2, 5 Mar 2026.
[3] Lin, G., Wan, Y., Pei, S., Xu, T., Xu, K., Xue, G. Behavioral Consistency and Transparency Analysis on Large Language Model API Gateways. IMC ‘26, Karlsruhe; arXiv:2604.21083v1, 22 Apr 2026.