Anthropic open-sources FlashProve to prove which AI model produced a given reply
Anthropic released FlashProve on 3 November 2026, an Apache 2.0 prototype that proves a reply came from a committed model without revealing its weights.
Summary
Anthropic released FlashProve on 3 November 2026, an open-source research prototype that lets an AI provider prove a reply came from a specific committed model while keeping the weights secret. Built by Anthropic's Frontier Red Team with EPFL, MIT and StarkWare, it runs Llama 4 Maverick on one NVIDIA H100 with a 1,024-token context. Anyone can check a proof on an ordinary CPU. It is Apache 2.0, unmaintained, and Anthropic has no deployment plans.