According to CoinMeta, Anthropic has open-sourced flashprove, applying zero-knowledge proof technology ( zksnark ) from the cryptocurrency field to AI model inference. This technology allows for the verification of blockchain transaction validity without disclosing transaction details, addressing the issue of users unsure whether service providers have switched to cheaper models when purchasing AI models API. flashprove enables service providers to generate encrypted fingerprints for model weights and provides mathematical proofs for inference results, allowing users to check if the output comes from the promised weights without needing to access the model weights themselves. Claude also participated in the development of this system. Anthropic has completed experiments on Meta's 400 billion-parameter model Llama 4 Maverick, processing proofs generated from 1024 token contexts, with each proof taking about 19MB. Ordinary CPU verification takes 17 seconds. Currently, each proof can only verify one output token; it is not yet possible to verify entire responses. The computational time required to generate a proof is 70–95 times that of ordinary inference. Whether it will be possible in the future to identify service providers that switch models using this technology depends on the reduction in proof costs.