Anthropic has released FlashProve, an open-source research prototype that attaches a cryptographic proof to an AI response so anyone can check which model actually produced it. Why it matters: today you have to take a provider’s word that the model you paid for is the model that answered — FlashProve is a first working attempt at making that claim verifiable instead of trusted.
November 3, 2026 — Anthropic published the research on Tuesday, and TechCrunch reported the release the same day.
What happened?
Anthropic’s researchers released a detailed technical specification and working code for FlashProve under an open-source license on GitHub. The system provides what researchers call “provable inference”: a mathematical guarantee that a given output came from a specific model. Anthropic is upfront that this is intermediate progress, not a finished system, and says there are no plans to deploy it at Anthropic or any other lab yet. The point of publishing it, the researchers say, is for others to take the work and improve it.
How does FlashProve work?
A lab first publishes a “fingerprint” — a commitment to one exact set of model weights. The weights themselves stay secret, which matters because closed-model makers treat them as their most valuable asset. When the model produces an output, the system generates a proof tying that output to the committed weights. Anyone holding the fingerprint, the prompt, and the reply can verify the proof, and Anthropic designed the verification to be lightweight enough to run on a local device.
TechCrunch compares it to cryptographic email signatures, which let a recipient confirm a message really came from its claimed sender. FlashProve is designed to hold up even against the lab itself: it could stop a company from quietly swapping in a cheaper model to cut costs, or a more expensive one to inflate benchmark results. It could also back up third-party claims — for example, that a software library really was scanned by a frontier cybersecurity model.
Why it matters
Model identity is a quiet trust problem that gets bigger as AI spending grows. Enterprises pay premium prices for specific frontier models, regulators are starting to ask which model did what, and right now neither group can check. FlashProve does not solve that today — it is a prototype with no deployment planned, and it still requires users to trust that a lab’s published fingerprint belongs to the model it names. But it shows the proof is technically possible without exposing the weights, which removes the main commercial excuse for not offering it.
Frequently asked questions
What is FlashProve?
An open-source research prototype from Anthropic that produces a cryptographic proof linking an AI response to a specific, previously committed set of model weights — without revealing those weights.
Can I use FlashProve with Claude today?
No. Anthropic says research is ongoing and there are no plans to deploy FlashProve at Anthropic or any other lab. The release is code and a technical specification for other researchers to build on.
What problem would provable inference solve?
It would let users verify they got the model they paid for, prevent providers from secretly swapping models, and let third parties prove claims that a specific advanced model performed a specific task.
Sources: TechCrunch, Anthropic