Anthropic open-sources FlashProve model-verification system
FlashProve uses zkSNARK proofs to confirm a response came from a given model without exposing its weights, Anthropic says.
Anthropic has released FlashProve, an open-source system that cryptographically verifies whether a given AI response was produced by a specified model. The announcement was reported by aitimes.com on November 2, 2026.
FlashProve relies on zkSNARK zero-knowledge proofs to confirm that a response matches a model's computation without revealing the model's weights. A service provider first publishes a digital fingerprint tied to the weights, then produces a proof file for each user request.
The check covers whether the output matches the specified model's computation, not whether the answer is factually correct or of high quality. In experiments, Anthropic used the 400 billion parameter mixture-of-experts model Llama 4 Maverick with public 4-bit weights and integer arithmetic. Proof generation ran on a single H100 GPU; each proof file was about 19MB, and verification took about 17 seconds using 24 CPU threads.
Anthropic reported that under certain experimental conditions, proof generation needed 70 to 95 times more GPU time than ordinary prefill with vLLM. The current implementation uses a 1024-token context and limited computation conditions, so it is not a finished product for commercial services with long conversations.
The researchers used the formal verification language Lean to check the correctness and security properties of the proof protocol, but did not prove the correctness of the entire implementation code. The technical report, proof generation and verification code, and Lean proof materials are available on GitHub under the Apache 2.0 license.
Why model attribution is hard
Until now, users could not easily confirm which model actually answered their request on an AI service, according to Anthropic.
Quick answers
What is FlashProve?
It is an open-source system from Anthropic that uses zkSNARK zero-knowledge proofs to verify that an AI response came from a specified model without revealing the model weights.
Does FlashProve check if an answer is correct?
No. It checks whether the output matches the specified model's computation, not whether the answer is factually correct or high quality.
Where is FlashProve available?
The technical report, proof generation and verification code, and Lean proof materials are published on GitHub under the Apache 2.0 license.