Claude, ChatGPT & Gemini1 source

FlashProve: Anthropic open-sources a way to prove which model replied

Anthropic released FlashProve, an open-source prototype (Apache 2.0). It uses zero-knowledge proofs to verify that a reply came from a specific model without revealing the weights. In the Llama 4 Maverick test it proves at about 110 tokens per second on one H100, but it costs 70 to 95 times more GPU and is limited to a 1,024-token context.

Why you should care

Right now you can't check whether an API serves the model you pay for or a cheaper one. This points to a future where that can be proven. It's research, not a product yet, but it shows where trust in AI APIs could go.

Written by my AI after reading 93 sources. I review it before it goes out (and cut the nonsense).

Sources

More from the radar

My community · opening soon on

Join the gang

The resources from every video ready to copy, challenges to actually do it and a place to ask without anyone giving you weird looks. Less theory, more tinkering.

Zero spam. I'll write when it's worth it and you can leave anytime.

Basic data protection information
Controller
Nic Ausió
Purpose
Letting you know when the community opens and sending you the weekly newsletter.
Legal basis
Your consent, which you can withdraw anytime.
Recipients
Resend (Resend, Inc., EE. UU.; con cláusulas contractuales tipo), which manages the list. No data is shared with third parties unless legally required.
Rights
Access, rectify, delete, object to, restrict and port your data by writing to hola@nicdopamina.com. You can also complain to the AEPD (Spanish data protection authority).
Full privacy policy (in Spanish)
  • Everything from the videos, at handPrompts, templates and flows ready to copy and paste.
  • Weekly challengesTo actually do it, not to bookmark it and forget it.
  • No-shame questionsNobody here knows it all. Me neither, relax.
  • You're in firstThe waitlist goes first. And meanwhile, every Thursday, the newsletter in your inbox.