Anthropic's FlashProve: proof of which model actually answered you

Anthropic published research on 'provable inference': FlashProve attaches a cryptographic proof to an answer showing it came from a specific, committed model, without revealing the model's weights. The demo runs on Llama 4 Maverick with a 1,024-token window at about 110 tokens a second on one H100, and costs roughly 70 to 95 times more GPU time than normal serving. The code is an open research release, not a maintained product.

Why it mattersNothing to use yet. Longer term, it could let you check that a provider really served the model you're paying for, which matters if you buy AI through resellers.

Curated by Sara Simeone as part of our human editorial coverage. The original sources are linked below; curation is distinct from a Hype or Fact? assessment.

SHARE THISPost on X ↗Share on LinkedIn ↗Copy link

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.