[Research: Anthropic Open-Sources FlashProve, Using Zero-Knowledge Proofs to Verify Large Model Inference] Anthropic has released an open-source research prototype, FlashProve, which uses zero-knowledge proofs to verify whether responses from large models are generated by specified model weights, without disclosing those weights. Anthropic states that Claude participated in improving the system; in tests on a 400 billion-parameter Llama 4 Maverick model, a single H100 GPU generated proofs at approximately 110 tokens (text units) per second, but the required GPU time was about 70 to 95 times that of ordinary inference. The prototype currently supports contexts of up to 1024 tokens and is still in the research phase.