Anthropic's FlashProve checks which model performed a computation

Anthropic has released FlashProve, an open-source research prototype for producing cryptographic evidence that a specified AI model performed a computation. Announced November 2, the system lets a provider commit to model weights and generate a proof that a client can check while the weights remain private. Its potential uses include checking the model behind a service and giving evaluators a stronger basis for attributing model outputs. Anthropic's research announcement

In the team's reported test, a single Nvidia H100 GPU generated proofs at roughly 110 tokens per second for a four-bit integer version of Llama 4 Maverick, with proofs around 19 megabytes per 1,024-token request. Proving used about 70–95 times the GPU time of the cited inference comparison. The prototype works after the fact and checks an integer model under greedy decoding, a precisely defined scope for the result. Technical report

The proof checks that the computation matches the committed model. Factual accuracy requires separate evaluation. Anthropic has released protocol documentation, code and mathematical security statements formalized in Lean; the repository identifies the software as a research prototype and marks it as unmaintained. Integrating this verification into real services will require greater efficiency and clear verification assumptions. FlashProve repository