AI Reports · Anthropic · 2026-11-02 · Blog / commentary
Progress on provable inference for large language models
Anthropic shares progress on zkSNARK-based cryptographic proofs for verifiable model inference, mainly through a system called FlashProve. The proofs let a model provider commit to a model and prove that each response came from it, which the post says is hard to check from outputs alone. According to the excerpt, for Llama 4 Maverick (a 400B-parameter mixture-of-experts model) with 4-bit quantized weights, a single H100 GPU generates proofs at about 110 tokens per second, with a proof size of about 19 MB per 1,024-token request.