Are you getting the AI you paid for? Anthropic’s new AI verification tool has a costly catch
Anthropic’s FlashProve research system demonstrates how users could verify that a specific AI model produced an output without revealing its weights. A provider commits to model weights, then supplies a cryptographic proof that the committed model generated the checked response; verification can run on a CPU. The demonstration uses a public, quantized version of Meta’s Llama 4 Maverick, not a feature available to Claude customers. Proof generation costs roughly 70–95 times ordinary inference in the reported comparison, and proofs are generated after responses, so they do not prevent users from first receiving altered outputs. The system is limited to 1,024-token contexts and exact greedy outputs, with unresolved implementation and security gaps. A trusted source must also identify the model commitment, and proof of computation does not establish that an answer is true or safe.
