Anthropic debuts FlashProve, a zkSNARK system that proves which AI model answered

Anthropic just pulled the curtain back on FlashProve, a zkSNARK system the company says can let AI providers cryptographically prove which model produced any given response. The company built the system with Claude, its own AI model, according to a report from Cointelegraph.

The pitch targets a quiet problem in AI: right now, when you ping an API, you largely take the provider's word for which model answered. FlashProve would swap that trust for a cryptographic receipt — a verifiable claim attached to each output.

zkSNARKs, or zero-knowledge succinct non-interactive arguments of knowledge, are the same family of cryptography that underpins privacy and scaling work across the crypto industry. In this case, the proof is aimed at model identity rather than hiding transaction data. That's a new twist on a tool crypto builders already know well.

The detail that raised eyebrows: Claude helped build FlashProve. Anthropic is using its own model to construct cryptography meant to keep AI providers honest — an AI writing the receipts for AI.

FlashProve lands as scrutiny of AI transparency grows, with users and enterprises increasingly asking for hard evidence about the systems behind the answers they get. The announcement, as reported, didn't include a rollout timeline or availability details.

For now, FlashProve is a claim about what's possible rather than a live product. If providers adopt it, verifying which model touched your data could stop being a matter of faith.