Anthropic has open-sourced FlashProve, a zkSNARK-based system intended to let users verify that an AI API output came from a committed set of model weights without accessing those weights, according to a Beating AI report carried by BlockBeats.
How FlashProve is described
Service providers generate cryptographic fingerprints for model weights and attach mathematical proofs to inference results. Users can then check the output against the committed weights, a design aimed at detecting an undisclosed switch to a cheaper model.
Current performance limits
The report said Anthropic tested FlashProve on Meta’s 400-billion-parameter Llama 4 Maverick. Proofs were about 19 MB per 1,024 context tokens, with standard CPU verification taking about 17 seconds. Each proof currently verifies one output token, while proof generation requires about 70–95 times the GPU compute of ordinary inference.
Source: BlockBeats English source, published Nov. 3, 2026.