Anthropic published FlashProve research on November 2, exploring cryptographic proofs that an AI response came from a committed model. The accompanying repository calls it a research prototype and says it is not maintained.
Identity is different from correctness
The aim is to help check whether a service used the model weights it claimed, rather than infer that from the style of an answer. A proof of the computation does not establish that the answer is true, useful, safe, or suitable for a particular task.
Anthropic reports proof-generation measurements for a quantized Llama 4 Maverick configuration, alongside substantial overhead compared with running inference alone. Those are the research team's measurements, not AiBrainWorX tests. The work is a prototype, and its practical deployment and costs need further evaluation.
Reported from linked primary sources. No product or model was independently tested.
