Anthropic Unveils FlashProve for Proving Which Model Answered You

What it claims

FlashProve, from Anthropic's Frontier Red Team, lets an AI provider prove that a reply came from a specific set of model weights without revealing the weights. The target problem is an API user who cannot tell whether the paid-for model answered or whether a router swapped in a cheaper one. On Llama 4 Maverick (400B parameters), the prover runs at about 110 tokens per second on one H100. Proofs are about 19 MB and verify on a plain CPU in about 17 to 21 seconds. The library is Apache-licensed.

How it works

The provider publishes a fingerprint of its weights once. It then attaches a proof to each reply showing that those exact weights produced it. The verifier needs only the published fingerprint, the reply and the proof, not the weights.

flowchart LR
  W[Private weights] --> F[Published fingerprint]
  W --> P[Per-reply proof]
  F --> V{CPU verifier}
  P --> V
  R[Reply] --> V
  V -->|match| A[Accept: model X said Y]
  V -->|mismatch| X[Reject]

The proofs are hash-based and belong to the STARK family used by Ethereum L2s such as Starknet. The report cites WHIR, STIR and the broader STARK research line. Lean-verified proofs ship alongside the code. Co-founders and fellows of StarkWare and Succinct are among the authors.

Limits

Proving costs about 70 to 95 times the GPU time of normal inference. Requests are batched, so a single request waits about five minutes for its batch. Context is capped at 1,024 tokens. Anthropic calls FlashProve a research prototype and is not maintaining the repository.

Onchain verification is not supported. Bankless notes that 19 MB proofs and 17-second checks rule out use in DeFi agents or prediction-market resolution without proof recursion, which is not part of this release.

This page is a Bankless news article. It does not give install commands, the repository location, the API shape, or which models besides Llama 4 Maverick were tested.