Anthropic and collaborators released FlashProve, an open-source zkSNARK-based prototype for provable inference that lets a model provider produce zero-knowledge proofs that a claimed response was generated by a committed model, enabling verifiable model integrity for users, auditors, and regulators. For a 400B-parameter Llama 4 Maverick (4-bit), they report ~110 tokens/sec proof generation on a single H100 and ~19 MB proofs per 1,024-token request, and ship code, a C++ verifier, and Lean formalizations under Apache 2.0.