# Anthropic open-sources FlashProve, a way to prove which model actually wrote a reply

> Labs would publish a fingerprint of each model, and anyone could check an output against it, without seeing the weights. Even the lab running the model could not fake it.

- URL: https://devisdead.ai/news/anthropic-open-sources-flashprove-a-way-to-prove-which-model-actually-wrote-a-re
- Published: 2026-11-04T03:30:52Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: anthropic, flashprove, cryptography, provable inference
- Originally reported by TechCrunch (https://techcrunch.com/2026/11/03/anthropic-is-working-on-a-system-to-prove-ai-models-identity/).

Every API call comes with an act of faith: you asked for the expensive model, and you hope that's the one that answered. Anthropic researchers want to replace that hope with math.

The system is called FlashProve, and as Russell Brandom reports for [TechCrunch](https://techcrunch.com/2026/11/03/anthropic-is-working-on-a-system-to-prove-ai-models-identity/), it targets so-called provable inference, meaning a cryptographic guarantee that a specific output came from a specific model. Think of it as a signed email, except the signature belongs to the model, and it is meant to hold up even if the impostor is the lab running it.

## What actually shipped

Not a product. Anthropic published a research post, a detailed technical spec and working code for parts of the system on GitHub under an open source license. The research isn't finished, and neither Anthropic nor any other lab plans to run it in production yet. The researchers call it intermediate progress and openly invite others to take it and improve it.

The design works like this:

- **Fingerprints:** a lab publishes a fingerprint for each model it releases. Outputs can later be matched to it and verified cryptographically.
- **Weights stay private:** the fingerprint doesn't reveal the model's weights, so closed labs could take part without handing over their most valuable asset.
- **Light checks:** verification is designed to be cheap enough to run on a local device.

## Why developers should care

The researchers' own examples are the right ones. A provider could quietly serve a cheaper model to save money, or run a pricier one to pad its benchmark results. A vendor could claim a frontier security model scanned its library, or that a business plan got a model's sign-off. Today you take those claims on trust. With proofs, you could check them.

That matters most for anyone building on resold or routed model access, where you often can't tell what's on the other end of the endpoint. It would also give audits and compliance reviews something firmer than a screenshot of a dashboard.

## The reality check

It only works if labs publish fingerprints, and nothing obliges them to. The idea has to win adoption across competitors who currently gain from being opaque, and a research post won't do that on its own. Proof also says which model answered, not whether the answer is right. Still, this is the kind of plumbing the industry usually builds only after a scandal, and it's nice to see it arrive first.

**[Scoreboard](/scoreboard) update:** Humans +1. Users may finally get a way to check that the model they paid for is the one that answered.

Stay undead.