Anthropic releases FlashProve open-source prototype for provable AI inference

Anthropic open-sourced FlashProve, a zkSNARK-based research prototype that generates cryptographic proofs verifying a response originated from committed model weights without revealing those weights. Demonstrated at approximately 110 tokens per second on a 400B-parameter model using one H100, it advances verifiable inference for trust, auditability, and safety evaluation in AI deployments.