Anthropic is developing an open-source verification system called FlashProve, which uses cryptography to provide mathematical guarantees for AI outputs—proving that they genuinely originate from a specified model and preventing impersonation between models. The system can close the cost-cheating loophole in which AI companies substitute cheap models for expensive ones, prevent fraudulent practices of artificially inflating model benchmark scores, and verify third-party claims involving model usage. Its key design is that AI labs only need to provide a public fingerprint for each model that does not reveal core weights, and users can complete verification locally. At present, the technical specification and runnable code have been open-sourced; research is still ongoing, and there are no plans to deploy it themselves for the time being.