Anthropic open-sources FlashProve to cryptographically verify which model answered

What happened

Anthropic has released FlashProve, an open-source prototype that lets an AI model provider commit to a specific model and then attach a cryptographic proof showing that each response was generated by it, Anthropic reported. The system relies on zkSNARKs, a class of zero-knowledge succinct proofs. The company said Claude produced the improvements that make it substantially faster than earlier open-source approaches.

In tests on Llama 4 Maverick, a 400B-parameter mixture-of-experts model using 4-bit quantized weights, one H100 GPU generated proofs at roughly 110 tokens per second. Each 1,024-token request yields a proof of about 19 MB. Proving costs around 70 to 95 times the GPU time needed to run the model itself on vLLM, an open-source library, while earlier open-source systems were described as orders of magnitude slower.

The proofs conceal the model's weights but are not designed to hide its architecture. To make Claude's design choices easier to audit, the team supplied Lean formal statements of the security guarantees, plus a prover, a fast C++ verifier and a provably correct Lean verifier. Code, proofs and documentation are available under the Apache 2.0 license, and the authors framed the work as intermediate progress.

Context

Users today cannot confirm which model produced a response. A routing service could quietly swap in a cheaper model, or an attacker could tamper with the token stream. Checking directly would require private weights and heavy compute. Anthropic said provable inference could also support independent evaluators and regulators, including proposals that would limit hardware to inference rather than training.