Anthropic says its FlashProve prototype can prove which model wrote an answer, at about 110 tokens a second on one H100

Anthropic's Frontier Red Team published progress on “provable inference” on 2 November. The provider publishes a commitment to a model's weights, then attaches a short zero-knowledge proof to each response, so that a user with little computing power and no access to the weights can check that the answer came from the model they asked for. Anthropic says this guards against a model being secretly swapped or its output tampered with.

The authors include researchers from Anthropic, EPFL, StarkWare, MIT and Succinct. Their system, FlashProve, was tested on Meta's Llama 4 Maverick, a 400-billion-parameter model, using its public 4-bit weights. Anthropic says one H100 produces proofs at about 110 tokens a second, and that this is about 70–95 times the GPU time of running the model on vLLM.

The page lists limits: proofs cover contexts of up to 1,024 tokens only, they cover an integer version of the model with greedy decoding, and side channels are out of scope. Anthropic calls the work a research prototype. It names possible uses, among them regulators checking that approved models are the ones being served, and customers checking they were not given a cheaper substitute.

• Claimed FlashProve proves Llama 4 Maverick inference at about 110 tokens a second on one H100, Anthropic says. Anthropic, Provable inference
• Confirmed Anthropic says proving costs about 70–95 times the GPU time of running the model itself on vLLM. Anthropic, Provable inference
• Confirmed The system currently proves inference over contexts of up to 1,024 tokens. Anthropic, Provable inference