Anthropic's Frontier Red Team, working with collaborators from EPFL, MIT and elsewhere, published FlashProve on November 2: an open-source research prototype for provable inference. Using a zkSNARK — a zero-knowledge proof system — a provider publishes a short commitment to one exact set of model weights without revealing them, then produces a proof for a completed prompt-and-response pair that anyone can check against that commitment.

The test case was large. Anthropic evaluated FlashProve on Meta's Llama 4 Maverick, a roughly 400-billion-parameter mixture-of-experts model in a 4-bit conversion, generating proofs at about 110 tokens per second on a single H100 GPU. A batch of 24 requests took about 223 seconds and produced an 18.78 MB proof per request, while verification took about 17 seconds on a 24-thread CPU with no GPU needed.

The cost is the catch. Proving took roughly 70 to 95 times the GPU time of simply running the model on vLLM, contexts are limited, and Anthropic frames the release — code, Lean proofs and documentation under Apache 2.0 — as a progress update, not a product. A proof also attests only that a specific model produced a response; it says nothing about whether that response was accurate, safe or unbiased.