[Anthropic Launches Zero-Knowledge Proof System for Verifiable AI Model Reasoning] According to TechFlow, on November 3rd, Anthropic released its research prototype, FlashProve, on November 2nd. FlashProve utilizes zero-knowledge proofs (zkSNARK) to achieve "provable reasoning," meaning a model provider can prove that a response was indeed generated by its claimed model without revealing the model's weights. The system first publishes a commitment to the model weights offline, then generates a proof for each request. Users can verify the proof in seconds on a regular CPU. For example, with the Llama 4 Maverick having 400 billion parameters, a single H100 can generate approximately 110 tokens of proof per second, and the proof for every 1024 tokens requested is approximately 19 MB. The system was optimized with the participation of Claude, and its security properties are formally described using Lean notation. Anthropic stated that the current proof cost is approximately 70–95 times that of conventional inference GPU time, and it remains a research prototype, not yet production-ready. Mert, CEO of Solana infrastructure company Helius, commented that Anthropic and Claude are using similar cryptographic techniques that Zcash pioneered globally.