Anthropic releases FlashProve to check AI outputs without revealing model weights
Research prototype published on Nov. 2
Anthropic has published FlashProve, a research prototype for checking whether selected AI outputs came from a committed model. The work was published on Nov. 2.
The demonstration uses Meta's Llama 4 Maverick model. Anthropic did not announce it as a verification feature for Claude customers.
How a commitment and a proof work
A provider first publishes a cryptographic commitment to its model weights. Model weights are the numerical parameters that determine how a model computes an answer. The commitment works like a fingerprint. It fixes which weights must be used, while keeping them hidden.
The provider then creates a proof that the committed model, given a specified input, produced the checked output. A verifier can check that proof without downloading the weights or running the model again.
A zero-knowledge proof is a method for checking computation without revealing the private information behind it. The source article says FlashProve uses the same broad family of techniques used in Ethereum layer-2 systems.
The commitment ID must come from a party the user trusts. It should not come from the server being checked or from files supplied with its proof. Without a trusted commitment, a proof does not show that the fingerprint belongs to the advertised model.
Key points
- FlashProve is a research prototype, not a deployed Claude feature.
- It checks model computation for checked outputs.
- It does not certify that answers are true, useful, or safe.
- Context is limited to 1,024 tokens.
- Implementation assurance remains limited.
Why a successful proof does not settle answer quality
A model can perform its own computation correctly and still give a factual error or unsafe advice. Checking the computation and assessing the answer are separate tasks.
FlashProve also creates proofs after a response by recomputing finished transcripts. This can detect tampering after delivery. It does not prevent a user from receiving an incorrect or altered response first.
Costs, speed, and present limits
Anthropic reports roughly 70 to 95 times the GPU time of ordinary inference in its benchmark. The comparison is per-GPU proving work against vLLM prefill using the FP8 version of the same model on eight H100 GPUs.
The demonstration uses a public 4-bit conversion of the 400-billion-parameter Llama 4 Maverick model. Proving runs on one H100 GPU. A user can check a proof on a CPU without a GPU.
The reported proof is about 19 MB per request. Verification takes 17 seconds on 24 CPU threads, or 21 seconds on one thread.
The reported 110 tokens per second counts all tokens processed in each 1,024-token request. In the main benchmark, each proof checks one following output token. The figure therefore describes processing throughput, not 110 individually authenticated new tokens per second.
The system proves exact integer greedy choices, meaning the highest-ranked next tokens. Ordinary floating-point computation or sampled responses can differ from those choices.
Anthropic provides formal security statements in Lean, a system for checking mathematical proofs. The fast executable code is not proved equivalent to that formal protocol. Zero knowledge has not yet been formally proved for batched requests. The repository labels the release an unmaintained research artifact.
CryptoSlate report on Anthropic's release
Newisty Editorial Team covers technology, cryptocurrency, digital products, online platforms, developer tools and the wider digital economy. Our content is researched from official sources, company announcements, public documentation, market data and other primary or reputable sources. Articles are reviewed and edited before publication for clarity, accuracy and useful context.
Related Posts
A $100 million crypto ETF launch balance doesn't prove real investor demand
Bitcoin Falls Below $84,000 as Markets Await U.S. Inflation Data
