Anthropic Releases FlashProve to Verify Which AI Model Answered

Anthropic, an AI research company, released a new open source tool called FlashProve on November 3. It lets you verify that a model's response is authentic, similar to checking a signature on a document. To make this work, Anthropic provides a "fingerprint" for each of its models. You can run these fingerprints on your own machine to confirm a response really came from the claimed model, without exposing its weights (parameters). The idea is to prevent two things: first, an AI lab quietly swapping in a cheaper model instead of the pricier one you asked for, and second, some outside party being able to tell which model was used for a given task (say, checking code for vulnerabilities).

This solves a real problem for anyone paying for models through an API: in theory, a provider could swap in a cheaper model without you ever knowing. I should note this is still just a prototype. No AI lab, including Anthropic itself, has built it into a product yet. Once the technology matures, expect this kind of check to show up in tools like Claude, where you'd be able to verify an agent's response the same way you now check its cost and performance.