When an AI service names the model behind an answer, customers normally have to trust that label. Anthropic’s open-source research prototype, FlashProve, explores a different approach: a cryptographic proof tied to a specific set of model weights.
A fingerprint and a proof
As Anthropic explains, a provider first publishes a short commitment—a fingerprint of the model’s weights, the numerical parameters learned during training. For a request, the provider then produces a separate proof. Someone with the commitment, prompt, output and proof can check whether the committed model produced those output tokens.
The fingerprint must itself be associated with the intended model through a trusted source. The proof addresses the computation’s origin, not whether an answer is true, safe or fair.
FlashProve uses a form of zero-knowledge proof: it is designed to let someone verify the claim without seeing the secret weights. The model architecture remains visible. Anthropic also says potential side channels, such as information revealed through timing, are outside this release’s scope.
What could this make possible?
A customer could have a way to check whether a provider substituted a cheaper model for the one promised. Safety evaluators could compare the model under test with the one served to customers. These are potential uses of the research, rather than features already offered across AI services.
The cost of checking
Anthropic evaluated the system with a 4-bit version of Meta’s Llama 4 Maverick. Its technical summary describes a proof of about 19 MB per 1,024-token request. Producing proofs requires additional GPU time compared with ordinary model execution, and the release remains a research prototype.
The code is public under Apache 2.0; its repository describes it as an unmaintained research prototype. It is not a universal authenticity badge automatically attached to Claude responses.
