Anthropic has released FlashProve, an open-source research prototype that produces cryptographic proof that a response came from a particular AI model. Published on 2 November by Anthropic’s Frontier Red Team with co-authors from EPFL, MIT, StarkWare and Succinct, it uses zero-knowledge proofs. The provider publishes a commitment to its model’s weights, then attaches a proof to each batch of answers. Anyone with the commitment, the prompt and the reply can check the proof without seeing the weights.
Why it matters: a customer paying for a model today has no reliable way to know they were not quietly served a cheaper or altered one. As Anthropic puts it, “judging from the model’s output alone is unreliable”.
The team demonstrated the system on Meta’s 400-billion-parameter Llama 4 Maverick. By Anthropic’s own measurements, proving ran at about 110 tokens a second on one H100 GPU, and checking a proof took about 17 seconds on an ordinary 24-thread CPU. The cost is steep: roughly 70 to 95 times the GPU time of serving the model normally.
There are limits. Proofs cover an integer version of the model with greedy decoding, context is capped at 1,024 tokens, and the fast verifier has no formal guarantees. Anthropic describes the work as intermediate progress, not a finished product. The code is on GitHub under the Apache 2.0 licence.