Anthropic's Frontier Red Team has released FlashProve, an open-source prototype that lets someone check that a reply came from a specific AI model. The research post is dated November 2. It is research, not a product, and the post says it is "intermediate progress, and not a completed research agenda."

What does FlashProve do?

Today you trust an AI provider's word about which model answered you. FlashProve uses a zero-knowledge proof (a zkSNARK) so a provider could instead prove it. It needs no trusted hardware, and Anthropic says the scheme offers post-quantum security.

The provider first publishes a commitment to the model's weights. Anthropic says that takes about 39 minutes once per model, and the commitment is about 6.5 MB and covers about 16 million requests. Each reply then gets a proof. Anyone holding the commitment, the prompt and the reply can verify it without the weights. The weights and the model's internal values stay hidden.

How slow is it?

Anthropic tested Llama 4 Maverick, a 400B-parameter model with 4-bit weights, on one H100 GPU with 24 CPU cores, for contexts up to 1,024 tokens. A batch of 24 requests took about 223 seconds. Each proof is about 18.78 MB, and checking one takes about 17 seconds on 24 CPU threads. Anthropic puts the cost at roughly 70 to 95 times the GPU time of running the model normally with vLLM.

What are the limits?

The proof covers an integer version of the model with greedy decoding, not the sampled, floating-point setup chatbots usually run. The model's architecture is not hidden. Timing side channels are out of scope, and it does not show the hardware isn't also doing other work. The fast C++ prover and verifier have no formal guarantees. The repository calls it "a research artifact: not maintained, not accepting contributions." The technical report has the details.

What this means for you

Nothing changes in any chatbot you use. FlashProve does not run on Claude, and Anthropic has announced no deployment. TechCrunch reports that a lab could publish a fingerprint of its model, which could help catch a quiet swap to a cheaper one. That is the idea, not a feature anyone offers. For where the industry stands on checking AI output today, see which chatbots watermark their text. For Anthropic's other recent news, see its updated IPO filing.

What we do not know

  • Whether Anthropic or any other lab will deploy this.
  • Whether it can reach sampled decoding, floating-point models, or longer contexts at a cost that makes sense.
  • Whether outside researchers will test the claims, since the code is unmaintained.