Anthropic, the company behind Claude, shared early research on a new open-source system called FlashProve. 

FlashProve aims to let people check which AI model produced a given output.

FlashProve

It is built around an idea researchers call “provable inference.” 

In plain terms, it creates a cryptographic proof that a specific model produced an output. The proof acts as a mathematical receipt. Nobody can fake it.

So if you think you’re talking to a top-tier model, FlashProve could confirm it. And if someone quietly swapped in a different model, the proof wouldn’t check out.

Anthropic compares the idea to digital email signatures. Those signatures let you confirm that a message really came from the person or company it claims to be from. 

FlashProve does something similar for AI answers.

FlashProve is designed to catch impostors, and that includes the AI lab itself. That might sound odd coming from an AI company. 

However, it addresses a major worry: a company could secretly switch to a cheaper model to save money. 

On the flip side, it could slip in a more powerful model during testing to make its results look better. Either way, users would have no idea.

With a system like FlashProve, those swaps would get much harder to hide. That’s a big deal for anyone paying for a specific model or comparing benchmark scores.

Checking Claims

Sometimes a third party claims that an advanced AI model did important work for them. A developer might say a top cybersecurity model scanned their software for weaknesses. 

Or a company might say a trusted model approved its business plan. In both cases, the researchers ask the obvious question: how would anyone check that?

FlashProve offers one possible answer. Instead of trusting a marketing line, people could check the proof themselves.

Fingerprints

For FlashProve to work, AI labs would publish a “fingerprint” for each model they release.

Anthropic’s research calls this a commitment. Later, an output can be matched to that fingerprint and checked using cryptography.

Importantly, the fingerprint doesn’t expose the model’s weights. Weights are the internal numbers that shape how a model thinks and responds. Companies like Anthropic and OpenAI guard them closely, since they’re among the most valuable things these labs own.

Because the fingerprint keeps those weights private, labs could share it without giving away their secrets. That means FlashProve could work with commercial, closed models, not just open ones.

Power

Anthropic also designed the checking step to be lightweight. You wouldn’t need a giant data center to confirm a proof. 

In fact, the research shows the checker runs on an ordinary computer processor, with no special graphics chip required.

In Anthropic’s tests, checking a single proof took around 17 to 21 seconds, depending on the setup.  The team had tested the system on Llama 4 Maverick, a large open model. 

Making the proofs is still costly, though. It takes far more computing power than simply running the model.

Research

Anthropic is clear that FlashProve isn’t finished; research is ongoing. Also, there are no current plans to roll it out at Anthropic or at any other lab.

The research page lists some of its limitations. For now, proofs only cover requests of up to 1,024 tokens, which is a fairly short chunk of text. 

The system also relies on a specific way of picking words, so some normal chatbot outputs may not match. 

Moreover, the fast code that runs the system doesn’t yet have formal guarantees behind it.

The researchers say they shared this work to report progress along the way, not to announce a finished product. They also openly invite others to study it, borrow from it, and make it better.

To support that invitation, Anthropic released a detailed technical write-up along with working code for parts of the system. 

Everything is on GitHub under the Apache 2.0 open source license. That means developers, researchers, and even rival labs can dig in.