Anthropic Launches FlashProve to Verify AI Model Outputs Using Zero-Knowledge Proofs

On November 2, 2026, Anthropic introduced FlashProve, an innovative system designed to enhance trust in AI model outputs by employing zero-knowledge proofs (zkSNARKs). This technology allows users to verify that the output generated by an AI model is indeed from the claimed model, rather than a potentially cheaper alternative.

Understanding FlashProve

FlashProve operates in two phases. Initially, during an offline step, the model provider commits to a specific set of neural network weights, which takes approximately 39 minutes to process and results in a commitment size of about 6.5 MB. For each inference request, the system generates a succinct proof that confirms the model produced the claimed output in response to a given prompt, all while keeping the model's weights confidential.

Performance and Challenges

Targeting mixture-of-experts (MoE) language models, FlashProve demonstrated impressive performance benchmarks on Meta’s Llama 4 Maverick model. However, the proving overhead is significant, running at about 70 to 95 times the GPU time required for standard inference. This makes FlashProve more of a proof of concept than a viable business model at present.

Development and Future Prospects

Led by Roy Rinberg and Alex Allain, the project features contributions from notable figures in applied cryptography. While the current implementation is not ready for production, it opens the door to future developments in secure AI model verification. The next challenge will be to reduce the overhead costs and proof sizes to make the technology practical for widespread use.