FlashProve is an open-source prototype for provable inference, published by Anthropic on 2 November 2026. It lets a user verify from a cryptographic proof, together with a commitment to the model weights, that a response came from the model a provider claims to be running, without the weights being revealed.
Benchmarked on Llama 4 Maverick, a 400 billion parameter mixture-of-experts model using roughly 17 billion parameters per token at 4-bit weights, proof generation reached about 110 tokens per second on a single H100 GPU and produced approximately 19 megabytes of proof per 1,024-token request. Proving cost roughly 70 to 95 times the GPU time of running the same model on vLLM.
The code, Lean proofs and documentation are released under the Apache 2.0 licence. Collaborators are drawn from Anthropic, EPFL, MIT, StarkWare, Succinct and RESI. Anthropic presents the work as a research prototype rather than a production service, and the overhead multiple is the reason why.