Anthropic open-sources FlashProve under Apache 2.0

Anthropic's FlashProve arrives against a specific threat model the company spells out: a user cannot check that a reply came from the model a developer claims is responding, as opposed to one secretly substituted by a routing service or an attacker who has tampered with the token stream. The system addresses that by committing to model weights and generating a zero-knowledge proof that the output tokens came from those weights. The guarantees of completeness, soundness and zero-knowledge were each formally verified in the Lean proof assistant. The release is tied to Anthropic's Frontier Safety Roadmap goal to develop a prototype by September 30, 2026 of provable inference, which the company says it met through a separate but related project. Independent coverage frames FlashProve as a check on inference vendors, but also flags the absence of an independent security audit and the 1,024-token ceiling as gaps before enterprise adoption. The prover's 72 to 94 times slowdown relative to a vLLM baseline keeps it an audit layer rather than a production inference path for most workloads.