Anthropic’s FlashProve aims to prove which AI model answered you

A lab’s weights, the numeric settings that make a model what it is, stay secret under this method, TechCrunch explains. What the lab has to publish is a fingerprint of every model, and a response can then be matched against it. The threat Anthropic names is a routing service, the middleman that forwards your request, quietly serving something cheaper to save money. The same check could back a vendor’s claim that a certain model reviewed a piece of work.

Speed is the catch for now. Anthropic reports about 110 tokens, the small chunks of text models work in, a second on one H100 graphics chip for a very large text model, with a proof file of roughly 19 MB per 1,024-token request. That file takes seconds to verify. The code is free under an Apache licence, the work covers text models, and TechCrunch says no lab, Anthropic included, has plans to deploy it.

Ad teams buy video models this way too. Higgsfield’s Starter plan, $19 a month for 270 credits at our 2 November check, sells other firms’ models such as Kling inside its own app (our guide to Kling on Higgsfield).

FlashProve covers text models and needs a lab’s fingerprint and a provider’s proof, so it gives a buyer paying a reseller for a named video model no check today, and that buyer is relying on the menu label.