AI
Anthropic ships an open prototype that proves which model answered
Anthropic published FlashProve, an Apache-2.0 prototype that lets a user check that a reply came from the model it claims to have come from. On Llama 4 Maverick, a 400-billion-parameter mixture of experts quantised to four bits, one H100 produced proofs at about 110 tokens per second, roughly 70 to 95 times the GPU time vLLM needed for the same work, with an 18.8 MB proof per 1,024-token request. Anthropic calls the result intermediate progress, not a finished product.