Anthropic releases FlashProve to cryptographically verify which model produced a reply

Anthropic released FlashProve, an Apache 2.0 zkSNARK (zero-knowledge proof) prototype that proves a reply came from a committed model without revealing the weights. It proves Llama 4 Maverick at about 110 tokens per second on one H100 GPU, with proofs around 19 MB.

Why it matters: Users of API resellers and third-party hosts currently have no way to know whether they get the model they pay for, so verifiable inference could curb silent model substitution. It also fits a broader push on provenance and trust, alongside the shared scraping opt-out standard and recent supply-chain incidents. Proof size and overhead still limit practical use for now.