I scrutinize privacy claims in AI products harder than almost anything else. The gap between "we protect your privacy" and what actually happens at the network layer is usually wide enough to drive a truck through.

Oblivious HTTP splits the knowledge of who you are from what you're asking. A relay knows your identity but not your prompt. The inference node sees your prompt but not your identity. OpenGradient applies this to its chat interface so no single party holds both pieces simultaneously.

What I'd verify is relay independence. If OpenGradient operates or controls the relay, the split is cosmetic. True anonymity requires a relay with no meaningful connection to the inference provider.

The architecture is sound. Relay ownership is what actually determines whether it holds.
#opg $OPG @OpenGradient
$SLX
$NES