Progress on provable inference for large language models

Anthropic has developed FlashProve, a cryptographic proof system that enables users to verify that responses from an AI model actually came from the specific model claimed, rather than a substituted or tampered version. The system uses zkSNARKs to generate proofs of inference, achieving approximately 110 tokens per second on a 400-billion-parameter model with ~19 MB proof sizes, and Anthropic is releasing the code, technical documentation, and formal Lean proofs under an open-source license. This advancement could support applications ranging from consumer security to AI compliance verification and helps address the inability of users to authenticate which model produced a given response.