Anthropic released FlashProve on November 2 to explore verifying that responses came from committed model weights. The proof hides those weights, but not the model architecture.

An experimental demonstration

The reported demonstration uses a quantized Llama 4 Maverick model, not Claude. Anthropic measured proof-generation GPU time at 70–95 times ordinary vLLM inference; this does not describe the model’s response speed.

Limits remain

The prototype supports windows of up to 1,024 tokens and proves an integer model’s greedy decoding. Each proof checks one token, not an entire 1,024-token answer. The Lean formalization does not guarantee equivalence with the fast executable code; side channels remain out of scope.

Code, proofs and documentation are Apache 2.0. Model weights are excluded and retain separate terms. The official repository calls this an unmaintained research artifact.