Anthropic introduces tool to verify AI model identity

Anthropic has released an open-source research prototype called FlashProve, a system designed to provide cryptographic verification for AI model responses. The tool aims to address transparency concerns by allowing users and auditors to confirm that a specific AI output was generated by a claimed model without requiring the disclosure of proprietary weights.

The system utilizes zero-knowledge proofs, specifically zkSNARKs, to establish this identity. According to Anthropic, the technology addresses a gap in the current AI landscape where users lack a technical method to verify the authenticity of the models they interact with, potentially preventing the substitution of cheaper or less capable alternatives.

Technical testing of the prototype involved a 400B-parameter mixture-of-experts model. On an NVIDIA H100 GPU, the system generated proofs at a rate of over 100 tokens per second. The resulting proof size is approximately 19 MB for every 1,024-token request, with verification taking about 17 seconds on a CPU using 24 threads.

The research team noted that the prover is currently 70 to 95 times slower than a standard vLLM baseline. Additionally, the system is limited to integer models using greedy decoding and supports contexts of up to 1,024 tokens. Anthropic developed this work in collaboration with researchers from EPFL, MIT, StarkWare, and Succinct.

The project was published on Nov. 2, 2026. Anthropic has categorized the release as a research prototype rather than a production-ready tool and has not announced a timeline for future deployment.