Anthropic open-sources FlashProve to verify AI model identities

Solving the AI trust gap

Many people in the technology industry are concerned about a growing lack of transparency in the AI market. There is a fear that some providers might secretly swap out expensive, high-end models for cheaper alternatives behind the scenes to reduce their operational costs. This practice could lead to lower-quality results for users who are paying for premium service tiers.

To address this specific trust issue, Anthropic has introduced an open-source prototype called FlashProve. The system is designed to function like a digital lie detector for artificial intelligence. It provides a mathematical guarantee that the answer a user receives actually originated from the specific model they requested.

By making this tool available to the public, Anthropic aims to create a standard for accountability. The project was developed in collaboration with researchers from several major institutions, including MIT, EPFL, StarkWare, and Succinct. This partnership ensures that the underlying technology is grounded in rigorous academic and engineering standards.

How zero-knowledge proofs work

FlashProve operates similarly to a digital signature on a secure email, but with a much more complex underlying mechanism. Before a lab releases a model to the public, they publish a unique digital fingerprint of that specific model. When a user asks a question, the server sends back the text answer along with a special cryptographic proof generated by the system.

The core of this technology is a security technique known as a zero-knowledge proof, or zkSNARK. This allows a third party to verify that a statement is true without the original party having to reveal any private information. In this context, it means a user can confirm which model generated a response without the company having to reveal its multi-million-dollar underlying code, known as model weights.

Because the system protects a lab's most valuable secret assets, it is highly attractive to commercial developers. They can prove their model's identity to customers while keeping their proprietary code completely hidden. This balance of transparency and privacy is what makes the tool viable for large-scale commercial use in the future.

Testing at massive scale

To demonstrate that the technology can handle real-world demands, the team tested FlashProve on a massive AI model. They used the Llama 4 Maverick model, which contains 400 billion parameters. This test served to show that the verification process could scale to handle large, complex models used by major tech companies.

The performance data shows that generating a proof for a 1,024-token request creates a file of approximately 19 MB. During the test, the system achieved a speed of 110 tokens per second when using a single powerful H100 graphics chip. While the AI company must expend significant computing power to create these proofs, the process is designed to be efficient for the provider.

For the end user, the process is remarkably fast and accessible. A user can perform a rapid validation on a typical desktop processor in roughly 17 seconds to confirm the AI's true identity. This disparity in effort means that while the heavy lifting happens on the server side, the verification remains a lightweight task for the consumer.

Open source and future uses

Anthropic has released all of the working code and technical details for FlashProve for free on GitHub. By using an open-source Apache 2.0 license, they are allowing anyone to study the mechanics of the system. This move is intended to push the entire industry forward toward more verifiable and trustworthy AI interactions.

Beyond simply catching companies that might be cutting costs, FlashProve has several other potential applications for businesses. For example, a developer could use the tool to confirm that their software underwent a security scan by a high-level AI. A startup could also use it to mathematically prove that an advanced AI approved its specific business plan.

However, it is important to note that this is currently an early research prototype. Anthropic has stated that it is limited to 1,024-token conversations and is not yet ready for wide-scale deployment. The company currently has no plans to integrate this into a live commercial product or a public API at this time.

Technical hurdles and costs

Despite its potential, there are significant technical hurdles that must be overcome before FlashProve becomes a standard feature. Generating these mathematical proofs is currently very expensive in terms of resources. The data shows that creating a proof requires a computational overhead of 72 to 95 times more than standard AI operations.

Because of this high computational cost, the tool is not yet practical for every single interaction a user has with a chatbot. It is currently best suited for specific verification tasks rather than every single turn of a conversation. The research team is likely looking for ways to optimize these proofs to make them more efficient for daily use.

For now, the project serves as a proof of concept for a more transparent AI future. It establishes a framework where mathematical certainty can replace blind trust. As the technology matures, it may eventually become a standard requirement for high-stakes business applications where knowing the exact model used is critical.