Anthropic is building FlashProve to verify which AI model answered

What Anthropic has announced

Anthropic is working on a system called FlashProve. According to TechCrunch, reporting on November 3, 2026, it would let anyone check that an AI response really came from the model it claims to be from. This article describes the stated goal, not a finished product, so we do not attribute technical parameters or effectiveness figures that the source does not provide.

For companies deploying AI, the issue goes beyond security. What matters is the ability to show which model produced a specific output in a specific process.

Why response provenance enters the audit

Companies usually rely on the provider's declaration of the model name and version returned by an API or platform. If that declaration cannot be checked independently, the event log records information nobody can confirm. In the EU, the reference point is Regulation (EU) 2024/1689 (AI Act). Article 12 requires high-risk systems to record events automatically in a way that makes their operation traceable, and Article 50 sets transparency obligations for interactions with AI systems. The regulation does not prescribe a specific method for verifying response origin, but in a compliance audit the question of which model actually answered becomes a practical part of these obligations.

How to prepare internal processes

Start with an inventory: which processes use external models, which version is declared, and which decisions based on AI output have legal or financial consequences. Then:

1. Log the model identifier returned by the provider together with the call time and prompt version.
2. Mark processes that require proof of origin, not just a declaration.
3. Define in the provider contract how model version changes are made and how you are notified.

The pitfall: a model identifier in an API response is a declaration, not evidence. Even if a provider offers a verification tool, it will not replace regression tests that check whether a version change altered results in your use case.

Where the risk lies

The most serious risk is a silent model swap or update that the team responsible for the output does not know about. Without an independent way to confirm response origin, it is hard to determine whether an error comes from the prompt, the data, or the model itself. The second risk is over-trusting a tool whose details are not yet publicly described in the source material. Before adding such a mechanism to a compliance process, review its documentation, scope, and the way it stores evidence.

Summary

Verifying that an AI response comes from the declared model could become a standard element of audits, especially in regulated processes. For now it is an announcement, so companies should already log model versions and agree on change rules with their providers.

Lub System helps B2B companies implement AI, automation and IT solutions end-to-end - from strategy to deployment. See our services or get in touch to discuss your case.