Anthropic reports progress on provable inference for large language models

Anthropic says FlashProve can generate proofs for Llama 4 Maverick at about 110 tokens per second on one H100 GPU. (Read More)Source