Anthropic's FlashProve prototype lets a client check which model produced an answer

Anthropic published a prototype for proving which model weights produced a response.

An AI application can ask for a model by name. Once a response comes back through a provider or router, the application generally has to trust that the named model served the request. On November 2, Anthropic published FlashProve, a research prototype that gives a verifier another way to establish the connection between a particular set of model weights and an answer. The release includes source code, a protocol description and a Lean formalization. [Anthropic research](https://www.anthropic.com/research/provable-inference) · [FlashProve repository](https://github.com/anthropics/FlashProve/)

How the proof ties an answer to a model

The provider first creates a cryptographic commitment to the weights. It can distribute this commitment to customers or auditors while keeping the weights private. For a response it wants to attest, the provider generates a proof. The client checks that proof against the commitment and the relevant input and output. The approach gives the client a way to verify the computation without acquiring a copy of the model or rerunning it. The protocol protects the weights while leaving the model architecture visible. [Anthropic research](https://www.anthropic.com/research/provable-inference)

A useful comparison is an auditor checking a model served through an intermediary. The auditor would need a trusted commitment for the agreed checkpoint, then could check whether the provider used those weights for the response under review. The commitment is part of the trust chain. A verifier needs a reason to associate it with the named checkpoint in the first place. The proof can attest to the computation while the evaluation still has to judge the quality of the answer.

What Anthropic measured

Anthropic demonstrated the implementation with a 4-bit version of Llama 4 Maverick, a mixture-of-experts model. Its reported test used one H100 for proving, a 1,024-token window and a proof that checked one token. The paper reports about 110 tokens per second for its batched proving measurement, an approximately 19 MB proof and 17 seconds for verification on 24 CPU threads. It also reports approximately 70–95 times the GPU time of ordinary inference in its comparison. These figures describe that workload and implementation. A reader should avoid treating proving throughput as the speed at which a user could receive a fully attested conversation. [Anthropic research and benchmark setup](https://www.anthropic.com/research/provable-inference)

Those costs point to a choice for prospective users. A team auditing a fixed benchmark or a sample of high-value responses may be able to budget for proof generation, transmission and storage separately from serving the model. Attesting every response would have a different cost profile. FlashProve gives teams public code to inspect and experiment with. The repository describes it as a research prototype. [FlashProve repository](https://github.com/anthropics/FlashProve/)