Anthropic Open-Sources FlashProve: Cryptographic Proof for AI Model Identity on Llama 4 Maverick

Anthropic announced on November 3, 2026, the open-sourcing of FlashProve, a prototype cryptographic system designed to verify the identity of the AI model responsible for generating a specific output. This development addresses a critical challenge in the AI landscape: the lack of reliable methods for users, evaluators, and regulators to confirm which AI model produced a given response. FlashProve leverages zero-knowledge proofs (zkSNARKs) to allow model providers to attach a cryptographic certificate to each AI output, confirming its origin from a specific, committed model without exposing the model's proprietary weights. For broader context, explore our AI Tools Pricing.

Addressing AI Model Verification Challenges

The increasing complexity and widespread use of AI models have highlighted a growing need for transparency and accountability. Without a verifiable method to confirm a model's identity, issues such as benchmark integrity, customer trust, and third-party verification claims become difficult to establish. FlashProve aims to resolve this by providing a mathematical proof of origin, ensuring that an AI output genuinely comes from the model it claims to. For broader context, explore our Top 100 AI Tools.

This system allows for model identity to be checked on standard hardware, which could significantly enhance trust in AI-generated content. For instance, it could help confirm that a specific version of a model was used for a particular task, which is crucial for regulatory compliance and maintaining consistent performance standards.

How FlashProve Works: Zero-Knowledge Proofs

FlashProve operates on the principle of provable inference, utilizing zero-knowledge proofs (zkSNARKs). In essence, a model provider commits to the weights of their AI model in an offline phase using a concise hiding commitment. Subsequently, for every client request, the provider generates a cryptographic proof that demonstrates the committed model correctly evaluated the client's input. This process confirms the model's identity and the integrity of its output without revealing the sensitive internal architecture or weights of the model itself.

The ability to hide model weights is a significant advantage, particularly for commercial entities that rely on proprietary models. This feature allows closed-weight laboratories to adopt FlashProve without compromising their intellectual property, fostering broader adoption of verifiable AI outputs.

Performance and Technical Specifications

Anthropic's research indicates specific performance metrics for FlashProve. The system generates proofs at approximately 110 tokens per second when running on a single H100 GPU. This performance was observed with a 400-billion-parameter mixture-of-experts model, specifically Llama 4 Maverick, which was 4-bit quantized.

However, the proving process introduces a notable overhead. The GPU time required for generating a proof is roughly 70 to 95 times greater than the GPU time needed to run the model itself on vLLM. Additionally, the proof size is approximately 19 MB for every 1,024-token request. These figures provide a clear picture of the computational resources and data transfer requirements associated with implementing FlashProve.

Open-Source Release and Future Implications

The release of FlashProve includes several key artifacts: a detailed research post, a comprehensive technical report, a protocol specification, executable code (including a prover, a fast C++ verifier, and a Lean verifier), and Lean proofs accompanied by an interactive visualizer. This open-source approach, under the Apache 2.0 license, encourages broader community engagement and further development.

This initiative aligns with a public goal outlined in Anthropic's Frontier Safety Roadmap, which aimed to prototype provable inference by September 30, 2026. While Anthropic views FlashProve as intermediate progress and does not have immediate deployment plans, this collaboration with institutions like EPFL, MIT, StarkWare, and Succinct underscores its significance in advancing AI safety and transparency. The availability of the code on GitHub allows developers and researchers to explore and build upon this foundational work.

Conclusion

Anthropic's open-sourcing of FlashProve marks a significant step toward establishing verifiable trust in AI systems. By providing a cryptographic method to prove the origin of AI-generated content, FlashProve addresses critical concerns regarding model identity, integrity, and accountability. While the system currently presents a considerable computational overhead, its open-source nature and the foundational research it represents lay the groundwork for future advancements in provable inference, potentially transforming how AI outputs are authenticated and trusted across various applications.

Sources

• Progress on provable inference for large language models \ Anthropic
• GitHub - anthropics/FlashProve: A research prototype for zero-knowledge proofs of AI inference. Not maintained. · GitHub
• FlashProve: an open-source library for provable inference
• Discovering cryptographic weaknesses with Claude \ Anthropic
• Anthropic is working on a system to prove AI models' identity | TechCrunch