Anthropic has released FlashProve, an open source prototype designed to cryptographically verify that a reply originated from a committed model without revealing the model’s internal weights. The system is built using zkSNARK technology, which generates lightweight proofs to confirm the source of AI outputs.
The prototype can verify inference from a 400 billion parameter Llama 4 Maverick model at a rate of 110 tokens per second on a single H100 GPU. Each verification proof is approximately 19 MB in size. Researchers published the technical specification and working code on GitHub under an Apache 2.0 license.
This tool aims to prevent AI companies from secretly swapping cheaper models to cut costs. It allows users to cryptographically confirm outputs from third parties, such as vulnerability scans. Anthropic states the research is ongoing and not yet ready for deployment, requiring labs to publish a fingerprint for each model.
