Anthropic’s Frontier Red Team has released FlashProve, an open-source cryptographic library designed to provide mathematical proof of an AI model’s identity. Released on November 2, 2026, the system aims to solve a fundamental problem in AI safety and regulation: how to verify that a specific response actually originated from a claimed model without requiring the provider to reveal proprietary model weights.
The system utilizes “provable inference,” a process where a model provider generates a cryptographic certificate alongside an AI’s output. This certificate allows a third party to verify that the text was generated by a specific version of a model, such as Claude 4 or Llama 4, using a “commitment ID” that serves as a digital fingerprint. This verification is powered by zero-knowledge succinct arguments of knowledge (zkSNARKs), which confirm the model’s identity while keeping the internal mechanics of the AI private.
The Technical Trade-off: High Overhead
While FlashProve represents a technical breakthrough in cryptographic verification, it currently carries a significant “compute tax” that prevents it from becoming a standard feature for consumer-facing AI. Generating these proofs requires between 70 and 95 times more GPU time than standard inference. Because of this overhead, the system is positioned as a research prototype for high-stakes scenarios—such as regulatory audits or sensitive government deployments—rather than a feature for everyday Claude users.
In technical demonstrations, Anthropic applied the system to Meta’s Llama 4 Maverick, a 400-billion parameter mixture-of-experts (MoE) model. According to Anthropic’s research report, the system achieved a proving speed of approximately 110 tokens per second on a single H100 GPU. For a typical 1,024-token request, the resulting cryptographic proof file is approximately 19 MB.
Collaboration and the Frontier Safety Roadmap
FlashProve was developed as part of a broad collaboration between Anthropic and researchers from MIT, EPFL, StarkWare, and Succinct. The project’s release follows a specific goal outlined in Anthropic’s Frontier Safety Roadmap, which called for a functional prototype of provable inference to be reached by September 30, 2026. While internal milestones were met earlier this fall, the open-source release allows the wider research community to inspect and build upon the codebase.
The current iteration of the software has specific technical limitations.
The development of FlashProve addresses a growing gap in the AI industry where model registries and regulators need a way to ensure that “safe” versions of models are the ones actually being served to the public. By providing a mathematical guarantee of identity, provable inference could eventually allow organizations to verify that a model has not been tampered with or replaced by a less secure alternative during deployment.
