FlashProve aims to prove which AI model really responds

Anthropic releases FlashProve, an open-source prototype that links an AI response to a specific model using cryptographic proofs. The tool remains limited and costly but targets a real blind spot in professional use cases.

Anthropic published on Tuesday, November 3, FlashProve, an open source prototype that aims to prove that an AI response indeed comes from a given model. The system relies on cryptographic proofs to link a result to a model fingerprint, without exposing proprietary weights. The issue is concrete for companies, auditors, and security teams: to know if the billed model is indeed the one responding. For now, the tool remains a lab research project, not a feature ready to plug into a product.

A cryptographic proof, not just a simple label

FlashProve is based on a simple idea to state, less simple to execute: the provider first publishes a cryptographic fingerprint linked to a specific set of weights, then generates a proof that connects a response to this model and the received prompt. This approach uses zkSNARKs, so-called zero-knowledge proofs that allow verifying a computation without revealing internal parameters. In short, one can verify the origin of a response without opening the black box.

Anthropic presents this mechanism as a form of "provable inference." The lab sees it as a tool for clients who pay for access to a specific model, but also for teams seeking to detect silent substitution. In practice, it could also serve to verify security audits or model comparisons. A less scrupulous provider would have less room to play magician.

What the prototype shows, and what it does not yet show

On paper, the idea addresses a real trust problem. Today, a provider can announce one model and serve another, less costly or more convenient to operate, without the end user easily noticing. Anthropic’s researchers also cite cases where third-party services presented Claude’s responses as coming from DeepSeek or Kimi. FlashProve makes this type of substitution detectable after the fact.

But the prototype has clear limits. During tests, Anthropic indicates having used Meta’s Llama 4 Maverick, with a throughput of about 110 tokens per second on an Nvidia H100 GPU. Each proof weighs about 19 MB and covers a single token within a 1,024-token context window. Verification takes 17 to 21 seconds on CPU, while proof generation consumes about 70 to 95 times more GPU time than a classic inference via vLLM. In other words, the system can count, but it cannot yet do so at commercial speed.

Why Anthropic is publishing this now

The timing is no coincidence. Anthropic says it wants to lay a technical foundation for future uses, without announcing deployment at its site or elsewhere. The code is published open source under the Apache 2.0 license, with the specification, so other teams can pick up the work. The company stresses the intermediate nature of the project: useful for research, not yet mature for a public or enterprise service.

This approach complements another ongoing effort around authenticating content generated by Claude. Here, it’s a different level: it’s no longer about marking that a text was written by AI, but identifying which model produced a given response. For API buyers, the nuance matters a lot. For the rest of the market, the promise is simple: less blind trust, more verifiable proof. And, incidentally, less room for cloud magicians.

Key points

  • FlashProve was published on November 3 by Anthropic.
  • The prototype is open source under the Apache 2.0 license.
  • Verification relies on zkSNARKs and a model fingerprint.
  • Tests cite Meta’s Llama 4 Maverick.
  • Anthropic does not plan any commercial deployment.

By the numbers

  • 1,024 tokens — prototype context limit, according to Anthropic.
  • 110 tokens/second — throughput observed with an Nvidia H100 GPU.
  • 19 MB — approximate size of a generated proof.
  • 17 to 21 seconds — verification time on CPU.
  • 70 to 95 times — GPU overhead compared to a vLLM inference.

To read