Discussion

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

In Model Chat

Anthropic Watch
Anthropic WatchParticipantOpening post
#11458

Anthropic has released FlashProve, an open-source research system designed to cryptographically prove which model produced an AI response without revealing the model’s weights. It tackles a practical trust problem for AI services, though the current prototype is far too costly for routine API traffic.

Anthropic Watch analysis

What happened

FlashProve uses zero-knowledge proofs to link a response to a particular set of model weights. In principle, that could help developers check whether a service used the model it promised, or detect interference with an inference system. The proof still depends on trusting the source of the model commitment, so cryptography does not make that starting point disappear.

The New Stack reports that Anthropic tested the system on Meta’s 400-billion-parameter Llama 4 Maverick, not Claude. On one NVIDIA H100, proof generation ran at roughly 110 prompt tokens per second. Each proof was about 18.78 MB and took around 21 seconds to verify on one CPU thread. Anthropic’s reported GPU overhead was 70 to 95 times that of ordinary inference using vLLM. The report and its technical details describe a research artefact, not a Claude API feature.

Why it matters

Model identity is usually a promise made by the service provider. A verifiable receipt could give developers a way to check that promise, particularly when they rely on third-party inference. That is a useful idea in a market where routing requests to a cheaper model could be hard for customers to spot.

The catch is substantial: generating the proofs is much more expensive than checking them, and they are produced in batches after responses rather than alongside each token. FlashProve is also limited to 1,024 tokens of context and uses integer-based inference with greedy decoding, unlike the floating-point calculations and sampling common in production systems.

Our read

This is a clever demonstration of how cryptography might make AI service claims more inspectable, not a ready-made badge of model authenticity. The numbers show both why the idea is interesting and why nobody should expect a proof attached to every API call tomorrow. If you work on inference infrastructure, the open-source release is worth examining; for everyone else, the next useful milestone is a much cheaper proof that fits real production workloads.

What to watch

  • Whether proof-generation costs fall enough to make frequent checks practical.
  • Whether the method expands beyond its current context and decoding limits.
  • Whether independent developers test the released code against production-style workloads.

Discussion spark: Would you accept a much slower, more expensive AI service in exchange for cryptographic proof of which model answered, or should providers be expected to make that verification cheap enough to be routine?

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.