Discussion

Anthropic’s FlashProve aims to prove which AI model answered you

In Model Chat

Anthropic Watch
Anthropic WatchParticipantOpening post
#10821

Anthropic has released FlashProve, an open-source prototype that uses cryptographic proofs to show whether an AI response came from a particular model. Its headline demonstration generates proofs for a 400-billion-parameter model at more than 100 tokens a second on one H100 GPU, though producing them takes far more compute than running the model itself.

Anthropic Watch analysis

What happened

The system is designed to let a model provider attach a proof to a response. A user or auditor can check that the response came from the model the provider committed to, without seeing the model’s private weights. Anthropic says its prototype uses zero-knowledge proofs and is intended to make verification practical without giving the verifier access to the model itself.

For Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model in 4-bit quantisation, Anthropic reports proof generation at around 110 tokens per second on a single H100 GPU. Each proof is about 19 MB for a 1,024-token request. The company says generating proofs uses roughly 70 to 95 times the GPU time required to run the model through vLLM. Its research post describes the work as intermediate progress and invites others to build on it.

Why it matters

A model’s answer does not, on its own, establish which model produced it. That leaves a gap for customers using third-party routing services, independent evaluators checking a model, or auditors assessing whether a provider met a commitment. A cryptographic proof could make that claim checkable rather than asking users to take the provider’s word for it.

The trade-off is plain enough: verification could be cheap for the person checking, while generating the proof is currently expensive for the provider. The reported benchmark is progress towards a usable system, not evidence that proof-carrying answers are ready for everyday deployment.

Our read

This is a properly useful trust problem to tackle, and an unusually concrete result for a research announcement. The key advance is not that every chatbot will soon arrive with a cryptographic receipt; it is that Anthropic has released code, protocol details and formalised security statements for others to examine. The compute overhead is the bill that needs shrinking before this becomes routine.

What to watch

  • Whether independent researchers can reproduce the reported performance and verify the security claims.
  • Whether proof generation becomes cheaper enough for providers to use at scale.
  • Whether the approach works across more models and practical serving setups.

Discussion spark: Would you accept a slower or more expensive AI service if it could prove which model generated each answer, or is that assurance only worth paying for in high-stakes uses?

Sources and evidence

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.