Anthropic releases FlashProve prototype for verifiable AI inference

What is changing?

Anthropic released FlashProve, a research prototype designed to verify that an AI response was produced by a specified model without exposing the model’s weights. The November 2 work aims to make it cryptographically auditable that a provider did not substitute a cheaper system or that an attacker did not tamper with the output stream.

In Anthropic’s reported test on Llama 4 Maverick, a single H100 GPU generated proofs at roughly 110 tokens per second using 4-bit quantized weights. The proof for a 1,024-token request was about 19 MB, while proof generation required roughly 70–95 times the GPU time of ordinary inference with vLLM.

Anthropic released a technical report, protocol description, C++ and Lean verifiers, executable code and Lean proofs for top-level security statements under Apache 2.0. FlashProve is not a finished product: the company describes it as intermediate research progress, and the performance results have not been independently validated in production environments.