Anthropic Says Its New Code Can Prove Which AI Model Answered You

The open-source FlashProve prototype checks each answer against a sealed fingerprint of the model, at roughly 70 to 95 times the GPU time.

Buy access to a top AI model through a middleman, and nothing in the reply proves it didn't come from a cheaper stand-in, Anthropic says. On Tuesday the company said it had a working prototype that could change that.

In a post on X, @AnthropicAI introduced FlashProve, free open-source code that lets developers confirm which model actually produced an answer. A research post dated Monday lays out the plan, and it's a receipt, not a promise. TechCrunch reported that neither Anthropic nor any other lab plans to deploy it.

Here's how it works. First the lab publishes a short fingerprint that locks in one exact set of model weights, the internal numbers that companies keep secret. After that, every answer can carry a proof file. Anyone holding the fingerprint, the prompt and the reply can check that file on an ordinary computer and know that the locked-in model produced exactly those words. Nothing about the weights leaks out.

Anthropic's post names who would want that. Outside testers who examine a model before and after launch need to know the model they're testing is the one customers get. A buyer whose requests get forwarded by a reseller wants proof the answer came from the model on the bill. Regulators who allow only approved models could ask for a proof with every response.

The cheater could even be the lab itself. TechCrunch reported that the researchers pitch the system as a way to catch a company that quietly swaps in a cheaper model to save money, or a pricier one to boost its test scores.

The trial ran on Meta's Llama 4 Maverick, a 400-billion-parameter model shrunk to 4-bit form from weights anyone can download. Locking in the fingerprint took about 39 minutes on a single Nvidia H100, and one fingerprint covers roughly 16 million requests before the lab has to make a new one.

Each 1,024-token request produced a proof of about 19 MB. Checking it took 17 seconds on a 24-thread processor, with no graphics chip.

The cost shows up on the proving side. One H100 proves about 110 tokens a second, which is a big jump. Measured against vLLM, a common open-source engine for running models, proving a batch of 24 requests took about 72 times the GPU time per chip. At the heaviest load Anthropic tried, it took about 94 times.

The proofs are also made afterward, in batches, so the guarantee shows up after the answer does.

Claude helped build the system, which created a problem of its own. Anthropic says Claude is strong at optimizing but makes choices that are hard to evaluate in a system this big. So the team wrote machine-checked proofs of the core security promises in Lean, a language for formal verification. According to Cryptopolitan, the GitHub page calls the code an unmaintained research artifact that hasn't had an independent security audit.

This also isn't Anthropic's only attempt. Its Frontier Safety Roadmap promised a provable-inference prototype by Sept. 30. A different project cleared that deadline, Anthropic says, by signing the company's own traffic and running the computation again on a server it trusts. FlashProve goes after the same goal with software alone.

By Wednesday the post had about 487,000 views on X, and crypto people were quick to claim the math as theirs. Helius founder Mert Mumtaz wrote on X that "anthropic/claude are now using the same cryptography that zcash was first in the world to pioneer," Cryptopolitan reported.

The overlap is a person. Alessandro Chiesa co-wrote the 2014 Zerocash paper behind the privacy coin Zcash. He's now at EPFL and is on FlashProve's author list. Math first used to keep payments private is now being used to show which model wrote a reply.

Anthropic calls the release intermediate progress and says it wants others to pick it up and improve it. Until some lab actually ships it, which model answered you is still the company's word.

Source: X, retrieved November 4, 2026. Other sources: anthropic.com, TechCrunch, cryptopolitan.com.