Anthropic Releases FlashProve to Verify AI Model Responses

Anthropic releases FlashProve, an open source prototype for verifying that an AI response came from a specified model, as detailed in a company research post. The system uses zkSNARK cryptographic proofs and does not reveal the model weights. Anthropic says Claude produced the improvements, building a system substantially faster than prior open source systems, and the team added Lean statements so the security claims can be audited.

In a test using Llama 4 Maverick with 400 billion parameters and 4 bit quantized weights, one H100 GPU generated proofs at about 110 tokens per second. Each 1,024 token request produced a proof of about 19 MB, while proof generation used about 70 to 95 times the GPU time required to run the model with vLLM.

Anthropic released the prover, a C++ verifier, a verifier written and checked in Lean, protocol documentation, and Lean proofs for the main security statements. The code, proofs, and documentation use the Apache 2.0 license.