Zcash (ZEC) fell about 5.9% in 24 hours to roughly $1,190 as the broader crypto market also declined. The drop came even as Helius founder Mert Mumtaz highlighted a fresh connection between Anthropic's newly released FlashProve tool and the privacy cryptography Zcash helped introduce to the world. Mumtaz's company builds infrastructure for the Solana blockchain, placing his commentary outside the Zcash project itself.

The pullback was market-wide: Bitcoin (BTC) traded near $76,673, and the total value of the crypto market fell about 1.8% over the same period.

Why Zcash is being linked to Anthropic's FlashProve

In a post on X published within the same 24-hour window, Mumtaz wrote that “anthropic/claude are now using the same cryptography that zcash was first in the world to pioneer,” ending with the phrase “probably nothing.” The post drew attention, gaining more than 1,000 likes and over 170 replies at the time of writing.

It was not Mumtaz's first time making the case for the privacy coin. In May, he appeared on the Bankless podcast and argued that Zcash is worth more than the market gives it credit for, especially compared with XRP and BNB. He said ZEC's value should represent a share of Bitcoin's total market value, with about 10% suggested.

Zcash's cryptographic lineage traces to the 2014 paper “Zerocash: Decentralized Payments from Bitcoin.” The paper's core tool, zk-SNARKs — shorthand for zero-knowledge succinct non-interactive arguments of knowledge — lets one party prove a statement is true without revealing the data behind it. That is the same property FlashProve now applies to AI inference: proving which model produced a response without revealing its weights. In 2024, the IEEE Symposium on Security and Privacy awarded the paper a Test of Time Award, with the committee stating it gave “an actual example for how to use zk-SNARKs in practice.” The Zcash network launched in October 2016.

One researcher is tied to both projects. Alessandro Chiesa co-wrote the Zerocash paper and is now at EPFL. He is also linked to StarkWare and appears on the FlashProve contributor list alongside Anthropic's Roy Rinberg and Alex Allain.

The overlap is not the first between Anthropic and Zcash this year. In June, Zcash founder Zooko Wilcox said an audit by Anthropic's Mythos AI, requested by Shielded Labs, found no serious new bugs in the protocol.

Notably, the Zcash community is currently arguing over its roughly $125 million grants treasury.

What FlashProve is, and what it costs to run

Anthropic released FlashProve, an open-source research prototype, on November 2. The tool uses a hash-based zkSNARK to prove that an answer came from a specific AI model without revealing the model's internal settings, known as weights. The project is documented on Anthropic's research page.

FlashProve is designed to close a trust gap. When a user sends a prompt to an AI service, they cannot be certain the reply came from the model they paid for: before an answer is delivered, a routing service could swap in a cheaper model, or the reply could be altered along the way. The tool is built for mixture-of-experts models, which route each input through a few specialized parts of the system.

FlashProve works in two steps. First, the provider publishes a small commitment of about 6.5 MB that locks in one exact set of weights while keeping them secret, a process that takes about 39 minutes. Second, for each prompt and reply, the system creates a short proof that the committed model produced those exact output tokens.

Anthropic tested FlashProve on Llama 4 Maverick, an AI model from Meta (NASDAQ: META). The model has 400.7 billion parameters and was run in a compressed 4-bit form. The test used a single NVIDIA (NASDAQ: NVDA) H100 chip, which handled about 110 tokens per second. A token is a small piece of text, such as a word or part of a word.

The system processed 24 requests in about 223 seconds. Each request was 1,024 tokens long and produced a proof of about 19 MB. Verifying a proof took only 17 to 21 seconds on a regular computer chip (CPU), and the checker did not need access to the model's weights.

The trade-off is speed: compared with vLLM inference, proof generation took about 72 to 94 times more time per GPU, depending on load.

Anthropic describes FlashProve as an intermediate research tool rather than production software. The GitHub page labels the code an unmaintained research artifact under the Apache 2.0 license, and the project states it has not undergone an independent security audit. For anyone tracking where the technology goes next, both the audit question and the ongoing treasury debate inside Zcash are the concrete items to watch.