Anthropic's Frontier Red Team released FlashProve, an Apache-licensed library that lets an AI provider cryptographically prove a response came from a specific set of model weights, without revealing the weights. It leans on the same family of hash-based proofs that power Ethereum L2s like Starknet.

— (@)

What's the Scoop?

• The problem: When you query an AI API, you can't tell whether the model you paid for actually answered, or whether a router quietly swapped in a cheaper one. FlashProve has the provider publish a fingerprint of its weights once, then attach a proof to each reply showing those exact weights produced it.

  Enjoying this article?

  Subscribe to Bankless or sign in

  Subscribe for free

• The numbers: On Llama 4 Maverick, a 400B-parameter model, the prover runs at about 110 tokens per second on a single H100. Proofs weigh in around 19 MB and a plain CPU can check one in roughly 17 to 21 seconds.

• The catch: Proving still costs roughly 70 to 95x the GPU time of normal inference, a single request waits around five minutes for its batch, and context is capped at 1,024 tokens. Anthropic calls it a research prototype and says it isn't maintaining the repo.

• Crypto's fingerprints: The report cites WHIR, STIR, and the broader STARK research lineage, and ships Lean-verified proofs alongside the code. StarkWare and Succinct, whose co-founders and fellows appear among the authors, were quick to claim a share of the win.

• Big picture: An agent or oracle that can prove "model X said Y" is a missing piece for AI-driven DeFi and prediction market resolutions. At 19 MB and 17 seconds per check, that's not happening onchain without recursion that isn't part of this release. But a frontier lab publicly building on the same math as Ethereum's ZK rollups is a legitimacy transfer the STARK camp has waited years for.