High-Tech and AI · Note
Anthropic open-sources FlashProve for cryptographically checked AI replies
A zkSNARK prototype that can prove a reply came from committed model weights, without revealing those weights.
Anthropic published FlashProve, an open-source research prototype for “provable inference”. The idea is a zkSNARK proof that a response came from a specific committed set of model weights, without handing those weights to the verifier.
On Llama 4 Maverick (4-bit), Anthropic reports proof generation around 110 tokens per second on one H100, with proofs of about 19 MB per 1,024-token request. This is not a Claude product switch you flip in the console; it is research code and Lean-audited claims aimed at future verification of model outputs.
Checked 7 November 2026. If the original page disagrees, follow that page.
Read the source Anthropic
