Anthropic releases FlashProve, an open-source prototype that cryptographically verifies an AI response came from a given model, not one that a router swapped in

You can now verify AI model responses were generated by the claimed model, not a substitute.

Anthropic has released FlashProve, an open-source prototype system designed to cryptographically verify that an AI model's response originates from a specific, intended model. This addresses the challenge of users being unable to confirm the identity of the AI model they are interacting with, as responses could potentially be substituted by routing services or malicious actors. FlashProve utilizes zkSNARKs (zero-knowledge succinct arguments) to generate proofs of inference. For a large 400B-parameter model (Llama 4 Maverick), the system can generate proofs at approximately 110 tokens per second on a single H100 GPU, with a proof size of about 19 MB per 1,024-token request. This performance is significantly faster than previous open-source systems. Anthropic is releasing the technical report, code, formal verification artifacts in Lean, and documentation under the Apache 2.0 license, encouraging community development and improvement.

FlashProve marks a significant step towards verifiable AI, introducing a practical mechanism for cryptographic proof of inference that promises to enhance trust and security in AI interactions. The performance metrics, demonstrating over 100 tokens/second proof generation for a large model, suggest that verifiable AI is moving from theoretical possibility to real-world application, potentially impacting how AI services are audited and regulated. This development could fundamentally alter the trust landscape for consumers and enterprises relying on third-party AI providers.

The open-sourcing of FlashProve, including its formal verification in Lean, invites scrutiny and collaborative improvement from the cryptography and AI research communities. Its application extends beyond simple user assurance, offering potential solutions for AI pacing regulations and tamper detection in sensitive AI deployments. As the technology matures, it could become a standard component for ensuring the integrity and compliance of AI systems, influencing both competitive dynamics among AI providers and the regulatory frameworks governing AI development.