Anthropic releases FlashProve to verify AI model responses

Anthropic has released FlashProve, an open-source research prototype that uses cryptographic proofs to verify whether a particular AI model generated a response, without revealing the model’s private weights. The company says the system is not planned for deployment and remains an early-stage project. FlashProve works by having a provider publish a cryptographic fingerprint, or commitment, tied to a model’s weights. A proof supplied with an output can then be checked against that fingerprint to establish that the committed model generated the response. The approach is intended to let users and evaluators check a model’s identity without relying solely on a provider’s assurance or requiring disclosure of its weights. Anthropic reported testing the prototype on Meta’s 400-billion-parameter Llama 4 Maverick, a mixture-of-experts model. On one Nvidia H100 GPU, it generated proofs at about 110 tokens per second. A proof for a request of 1,024 tokens was about 19 megabytes and took roughly 17 seconds to verify on a CPU. Anthropic said proof generation used about 70 to 95 times the GPU time of running the model normally. Those results illustrate both the prototype’s technical...