In early November 2026 Anthropic published progress on provable inference and released FlashProve as an Apache 2.0 research prototype. The system uses zkSNARKs to prove a reply came from committed model weights without revealing them. On Llama 4 Maverick (400B MoE), one H100 generated proofs at about 110 tokens per second with ~19 MB proofs per 1,024-token request. The release includes C++ and Lean verifiers.

What it means: Proving which model produced a reply matters for trust in third-party routers, safety evaluations and possible regulation. Greek banks, hospitals or public bodies buying APIs through resellers could eventually cut model-substitution risk. FlashProve is still a research prototype with limits such as a 1,024-token window, an integer model and no product support. It does not change procurement contracts today, but signals a path toward verifiable AI compliance. For Greek businesses and public bodies this means a practical review of contracts, timelines and compliance duties, and it is not investment or legal advice. IT, finance and compliance teams should document owners, review cycles and fallback processes before wider adoption or a vendor change. For Greek businesses and public bodies this means a practical review of contracts, timelines and compliance duties, and it is not investment or legal advice. IT, finance and compliance teams should document owners, review cycles and fallback processes before wider adoption or a vendor change.

Read the original on Anthropic

Related service: Custom software

Related stories