Is the model you paid for really the one answering you? Meet FlashProve
When we use a large language model API, we blindly trust the service provider. We assume that if we paid for an expensive flagship model, that is the model actually processing our requests. But do we have a way to verify this? Today, service providers can swap models "under the hood"—using a smaller, cheaper model instead of the promised one, or even changing the model's behavior without our knowledge.
This is where a new technological breakthrough from Anthropic called FlashProve comes into the picture [1, 3].
What is FlashProve?
FlashProve is an open-source system that enables "Provable Inference" [1, 2]. The idea is simple yet powerful: by using advanced cryptographic techniques called zkSNARKs (zero-knowledge proofs), the model provider can generate a mathematical proof that the output received was indeed generated by the specific model weights they committed to [5, 8].
The beauty of this method is that it allows the user to verify the result without gaining access to the model weights themselves, thereby preserving the intellectual property of the AI companies [2, 3].
How does it work in practice?
The process consists of two main stages:
- •The Commitment stage: The provider publishes a cryptographic "fingerprint" of the model weights. This is a small piece of data that any client can store [1, 8].
- •The Proof stage: For every request, the provider generates a mathematical proof that links the input (the prompt) to the output (the generated tokens), using the previously published commitment [2, 9].
In an experiment conducted by Anthropic on the Llama 4 Maverick model (a Mixture-of-Experts model with 400 billion parameters), the system was able to generate proofs at a rate of approximately 110 tokens per second on a single H100 GPU [2, 7].
Are we there yet?
It is important to be realistic: this is a research breakthrough, not a mass-market off-the-shelf product. Currently, generating the proof requires significant computational resources—70 to 95 times more GPU time compared to running the model itself without a proof [5, 10]. Additionally, the size of the proof for a 1,024-token request is about 19 megabytes [2, 7].
However, the ability to verify the model's identity is critical for:
- •Data security: Preventing impersonation by malicious models.
- •Auditing and compliance: Ensuring that service providers meet their commitments.
- •Trust in autonomous systems: Ensuring that AI agents use the correct model before making critical decisions [3, 9].
Summary and conclusions
FlashProve marks the beginning of a new era where "Zero Trust" becomes the standard in the AI world as well [4, 6]. While the computational cost is still high, the fact that Anthropic released the code and documentation under an Apache 2.0 license [5, 7] will undoubtedly accelerate research in this field.
We are still far from the day when every response from ChatGPT or Claude will come with an attached cryptographic proof, but the first step is already here. For those who want to dive deeper into the technology, you can find the code and technical reports on the project's GitHub [3, 4].
Written by Daniel Rechtman – Expert in Cloud, Artificial Intelligence, and No-Code development. Telegram | WhatsApp | Instagram