Anthropic's FlashProve proves which AI model actually answered
FlashProve is an open-source zkSNARK prototype that cryptographically proves a reply came from specific committed model weights, at roughly 70 to 95 times the GPU cost of running the model itself.
Anthropic released FlashProve on 2 November, an open-source prototype that lets a provider cryptographically prove a response came from a specific set of model weights, without revealing those weights.
It works by committing to a model once, then generating a zkSNARK-style proof for each inference that a verifier can check without any GPU. Anthropic tested it on Llama 4 Maverick, a 400B-parameter model with about 17B active per token, using a public 4-bit conversion. A single H100 generates proofs at around 110 tokens a second, which Anthropic says is roughly 70 to 95 times the GPU time of just running the model on vLLM. A verifier can then check a proof in 17 seconds on ordinary CPU threads. The system currently only proves inference over a 1,024-token context window.
Why it matters: if you buy model access through a router or reseller, you currently have to take their word for which model actually answered. FlashProve points at a way to verify that cryptographically instead, alongside uses like proving a safety-vetted model carried out a given analysis.
The caveat: Anthropic labels the code a research artifact, not maintained and not accepting contributions, and the overhead and tiny context window put it well short of anything you would run in production today.