FlashProve is Anthropic's open-source research prototype for showing, with a proof, which AI model produced a reply, without exposing that model's weights.
In plain words
Anthropic warns that a stronger model might secretly replace an approved one and pretend to be it. FlashProve makes that checkable. For every reply, the provider's GPU builds a zero-knowledge proof tying the reply to a model the provider has committed to, and anyone can check it on an ordinary processor in seconds without the weights. Published in November 2026, it proves about 110 tokens a second for Llama 4 Maverick on one NVIDIA H100, but at roughly 70 to 95 times the GPU time of just running the model.
A quick example
Imagine a hospital that pays for a specific, safety-tested model. Each reply arrives with a proof, and the hospital's own computer checks it, so a cheaper or different model can't quietly be swapped in.
Where you'll meet it
- Checking that an AI provider really served the model it promised.
- Making it harder to secretly swap a different model in for an approved one.
Related
Sources
- DocsProvable inference: FlashProve, Anthropic · read 3 Nov 2026
- Docsanthropics/FlashProve, GitHub (Anthropic) · read 3 Nov 2026
Checked by WhatIsAI against the sources above · 3 Nov 2026. Spotted an error? Report it.