Anthropic’s Frontier Red Team released FlashProve on 2 November 2026, an open-source research prototype designed to show cryptographically that a particular language model produced a particular answer. Researchers from EPFL, MIT and StarkWare collaborated on the project.
The release addresses a basic trust problem for hosted AI: a convincing answer alone does not establish which model generated it. A routing service could substitute a cheaper model, or someone with access to the inference server could alter the token stream, without the change being obvious to the customer.
FlashProve uses a zkSNARK-based approach. In an offline phase, the provider publishes a short commitment, effectively a fingerprint, to an exact set of model weights. For a prompt and its output, the provider can then create a proof tied to that commitment. A customer can check the proof without receiving the weights or running the model; the zero-knowledge design keeps the weights hidden, though not the model architecture.
Anthropic tested the prototype with a 4-bit version of Meta’s 400-billion-parameter Llama 4 Maverick mixture-of-experts model. It reports about 110 tokens per second on one H100 GPU. A proof for a 1,024-token request is about 18.78 MB and took roughly 17 seconds to verify on 24 CPU threads. Generating the proofs required about 70–95 times the GPU time of ordinary inference in the reported comparisons with vLLM.
The repository contains the prover, C++ and Lean verifiers, formal Lean proofs and a protocol specification. Code, proofs and documentation are released under Apache 2.0.
This is not yet a production safeguard for Claude or other hosted services. TechCrunch reports that there are no current deployment plans. Anthropic also lists substantial limitations: contexts are capped at 1,024 tokens, the proof covers an integer model using greedy decoding, the current design targets mixture-of-experts models, and the team does not guarantee that the paper, Lean protocol and fast executable code all match exactly. FlashProve therefore demonstrates progress toward verifiable model identity rather than a finished guarantee for everyday chatbot traffic.