Progress on provable inference for large language models

Anthropic researchers reported progress on cryptographic proofs that a specific AI model produced a given response, unveiling a system called FlashProve that generates zkSNARK proofs for inference on Llama 4 Maverick, a 400B-parameter mixture-of-experts model. The work, involving collaborators from EPFL, MIT, StarkWare, Succinct and Anthropic's fellows programs, aims to let providers commit to a model and prove each response came from it, addressing the risk that a routing service or attacker silently substitutes another model. Anthropic said prior open-source systems are orders of magnitude slower, and it is backing its security claims with top-level auditable statements written in the Lean formal verification language. The release includes five artifacts: the blog post, a technical report, a Claude-written protocol specification, and executable code containing a prover and two verifiers, including a fast verifier in C++.