Anthropic has released FlashProve, a research prototype for proving that an AI response was generated using a particular committed model. Published on November 2, the work targets a question that ordinary output inspection cannot reliably answer: did the service actually run the model it promised?
The research announcement describes zero-knowledge proofs that allow verification without revealing the model’s weights. Its demonstration uses a four-bit version of the 400-billion-parameter Llama 4 Maverick mixture-of-experts model. Anthropic reports proof generation at roughly 110 tokens per second on one H100 GPU, with approximately 19 MB of proof data for a 1,024-token request.
The overhead remains substantial: Anthropic reports about 70–95 times the GPU time of running the corresponding inference with vLLM. The result is intermediate research progress, rather than an announcement that ordinary Claude responses now arrive with cryptographic certificates.
Artifacts that others can inspect
The public repository contains the research implementation, protocol documentation and formal-verification material under Apache 2.0. It includes a prover and alternative verifiers. The repository describes the project as a research prototype and states that it is not maintained, which matters for anyone considering operational use.
The intended guarantee concerns execution against a mathematical specification. It does not establish that an answer is factually true, that the training data were appropriate or that the model is suitable for a particular decision. Those remain separate questions.
Where the guarantee could help
Our analysis is that inference integrity becomes especially useful when multiple organizations sit between a user and a model. A company may buy access through a routing service, while an evaluator may need confidence that the deployed system matches the one examined before release. A verifiable commitment could make those relationships easier to audit.
Turning that idea into a usable service would require careful handling of identity and updates. Someone must establish which commitment corresponds to the model a customer selected. A provider would also need a clear process for announcing a new checkpoint, retaining evidence and explaining which parts of the computation a proof covers.
Performance is another practical constraint. The cost of generating, transmitting and checking evidence must fit the application. A high-assurance evaluation may tolerate overhead that would be unacceptable in a high-volume consumer interaction. That suggests different deployment choices even if the underlying cryptography is sound.
FlashProve gives researchers a concrete system to examine and improve. Its immediate contribution is a public implementation and formal material around a narrowly defined trust claim. The longer-term question is how far the overhead can fall while preserving a guarantee that independent users can meaningfully verify.