Anthropic releases FlashProve to show which AI model gave an answer

Anthropic and university researchers released FlashProve, an open source system that shows which AI model wrote an answer. The model provider adds a proof to each answer, and one H100 GPU makes proofs at 110 tokens (small pieces of text) each second. The test model has 400 billion parameters, and each proof is about 19 megabytes for 1,024 tokens.

How it works

The provider adds a short proof, a zkSNARK, to each answer, and the user can examine it quickly. Anthropic tells that Claude increased the speed of the system, but its work is hard to check. The team gives top level statements in Lean, a language for formal proof, to help an audit.

The numbers

The test used Llama 4 Maverick with 4 bit weights. The proof time is 70 to 95 times the GPU time of vLLM, an open source library. Anthropic tells that other open source systems use much more time.

What is not known

Anthropic tells that this is first work and not the end of the research. The proof does not show the weights, but the structure of the model is open. The test used one model, and Anthropic gives no cost for a user.

Insight: A proof of identity does not show that a model is safe

In October 2026, a study put an encrypted network in the weights of a model, and audits did not find it. A user cannot know if a router put in a different model, Anthropic tells. A proof connects an answer to a selected model, and trust moves from tests to the selection of the provider. For leaders, a proof tells who gave the answer, not the content of the model.