Anthropic Releases FlashProve for Cryptographic AI Inference Proofs

Anthropic launched FlashProve, an open-source prototype that uses cryptographic proofs to verify that a language model's output originates from committed weights without revealing them. It achieves proof generation at 110 tokens per second on Meta's Llama 4 Maverick, enhancing trust in hosted AI services by enabling verification without rerunning models or exposing proprietary weights.