Back to collectionFlashProve Makes Model Identity Testable, at a Substantial Cost
Anthropic’s open research prototype checks computation against committed model weights; it does not establish that an answer is true.
Anthropic released FlashProve on November 2: an open-source research prototype that lets a verifier check whether output matches a committed set of model weights without receiving those weights. Its reported Llama 4 Maverick benchmark reaches about 110 proved tokens per second on one H100, but requires roughly 70–95 times the GPU time of ordinary inference. This makes model substitution more testable, rather than making verification free.
The present guarantee concerns an integer model with greedy decoding and contexts of at most 1,024 tokens. It is not a proof of an answer’s factual accuracy or a guarantee for arbitrary sampled chat responses. Anthropic also says its fast executable code lacks the formal guarantees attached to its Lean protocol and slower verifier. Those limits matter before the prototype is treated as production assurance.