Anthropic proposes FlashProve, using zkSNARKs to prove a response came from a specified model

Tue, Nov 3 15:49 UTC+8Updated Tue, Nov 3 16:45 UTC+8

Anthropic's research prototype FlashProve uses zkSNARKs to prove a reply came from a specified model. On 4-bit Llama 4 Maverick (400B MoE), one H100 proves about 110 prefill tokens per second at a batch size of 24, and each proof checks a single output token. Proofs are about 19 MB per 1,024-token request, at 70–95x vLLM's GPU time. Open source under Apache 2.0.

Why it matters

Users and evaluators can't currently confirm a response came from the claimed model; this makes model swaps by routers and tampered token streams detectable.

Sources1

AnthropicOfficial · WebsiteToday 15:49 UTC+8Progress on provable inference for large language modelsView original · Today 15:49 UTC+8

Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.

Long reads on this

Anthropic's FlashProve can prove which model wrote a response. Here is how far the research prototype gets5 min