Anthropic released FlashProve, an open-source prototype that attaches a cryptographic proof to an AI model’s response showing which set of weights produced it, without exposing the weights themselves. The proofs could let routing-service customers confirm they got the model they paid for, let regulators check that providers serve only approved models, and catch servers where an attacker has swapped in different weights. On Meta’s 400B-parameter Llama 4 Maverick, a single H100 GPU generated proofs at about 110 tokens per second, using 70 to 95 times as much GPU time as running the model normally, and each 19 MB proof checks in seconds on a regular CPU. Anthropic said Claude built much of FlashProve, which is still a research prototype limited to 1,024 tokens of context.
