Anthropic Unveils FlashProve for Proving Which Model Answered You

What it is

FlashProve is an Apache-licensed research library from Anthropic's Frontier Red Team. It lets an AI provider prove that a response came from a specific set of model weights without revealing them. The problem it targets is a router quietly swapping a cheaper model in behind an API.

How it works

The provider publishes a fingerprint of its weights once. Each reply then carries a proof that those exact weights produced it. The proofs are hash-based, from the same family used by Ethereum L2s such as Starknet. The report cites WHIR, STIR and the wider STARK research lineage, and the code ships with Lean-verified proofs. The page says StarkWare and Succinct co-founders and fellows appear among the authors.

Preconditions / Affected

The figures come from Llama 4 Maverick (400B parameters) on a single H100:

• Proving runs at about 110 tokens per second.
• Proofs are about 19 MB.
• A plain CPU checks one in roughly 17 to 21 seconds.
• Proving costs roughly 70 to 95x the GPU time of normal inference.
• A single request waits about five minutes for its batch.
• Context is capped at 1,024 tokens.

The page reports no results for other models or hardware.

Caveats

Anthropic calls it a research prototype and says it is not maintaining the repo. The page's own assessment is that onchain verification of AI outputs, such as DeFi agents or prediction market resolution, is not feasible at 19 MB and 17 seconds per check. It would need recursion, which this release does not include. The five-minute batch wait and the 1,024-token cap also rule out interactive or long-context use as released. The page is a news summary and does not describe the proof construction in detail.