zkAPI mainnet lets a metered API bill a proof, not a name

The zkAPI mainnet is a payment trick for metered APIs, and the first use the authors name is AI. You deposit credits once on Ethereum. After that, a proof says a funded note covers the spend, without telling the checker which note is yours. The Foundation post and The Block's October 1 story agree on that shape. They do not print the same calendar for the design paper underneath it. This is not a product review and not a trade.

TL;DR - The Ethereum Foundation post, by Vittorio Rivabella of the dAI team and dated October 1, 2026, says zkAPI is live on Ethereum mainnet. The Open Anonymity Project built it with the Foundation. - A deposit of ETH, USDC, or similar becomes a private note. Spending publishes a nullifier so the same note cannot be spent twice. The client speaks the OpenAI and Ollama APIs on localhost. - The Block, timestamped October 1, 2026, at 8:22 p.m. EDT, adds a quote from Ken Liu and says the design post on Ethereum Research is from Feb. 11. The Foundation background section names the authors and does not print that day.

What actually happened

The post's problem statement is ordinary and sharp. An API key points at an account, the account points at a payment method, and every prompt joins that record. The provider can stitch years of use into one profile. Prompts are personal. People ask models about health, money, and doubts. The post says the old alternatives were poor: pay per request onchain, which is slow, expensive, and traceable, or trust a middleman not to look.

On zkAPI mainnet, the mechanism is two paths that share a short-lived key and a usage receipt. You deposit into a vault contract. The balance lives as a commitment in a Merkle tree, so a proof can say "this note is one of the valid ones" without pointing at it. Each spend publishes a nullifier, a one-way serial number from the note's secret. Stay inside the balance and the spends do not link. Try to spend the same balance twice and the duplicate nullifier exposes the attempt and, the post says, nothing else.

The spending cap is a reservation. When the key expires, the provider side records total usage in a signed receipt, and the server deducts that amount from the note. One authorization can cover a session, not one proof per call. The post also describes a simpler proxy mode, where the zkAPI server relays requests to the provider. Easier to run, and the relay sees traffic. The runtime-key mode exists so the payment intermediary does not.

For the zkAPI mainnet, the client, the post says, exposes the standard OpenAI and Ollama APIs on your own machine. Existing apps point at localhost. The first use case is AI inference, because prompts are sensitive. The same post lists other metered jobs: blockchain RPC queries, image and video generation, VPN bandwidth, and machine-to-machine payments so agents can pay without an account. The vault link on the post is ZkApiVault on Ethereum mainnet, described as holding USDC credits. There is also a Sepolia deployment and an OA Chat demo. Those are links on the Foundation page, not a usage statistic.

What The Block adds

The Block's zkAPI mainnet story is by Zack Abrams, edited by Timmy Shen, and the structured date is October 1, 2026, 8:22 p.m. EDT, updated 8:24 p.m. EDT. That is 3:22 a.m. Moscow on October 2 for the first timestamp. It says the Foundation launched the system on Thursday, which matches October 1, 2026. It quotes the post's line about every AI API call carrying an identity. Then it adds a source the post does not: Ken Liu, described as a Stanford computer science PhD candidate who works on the Open Anonymity Project, posting on X on Thursday afternoon. The quote The Block uses is that the system is like a generalization of OA unlinkable inference that also abstracts payments away, and that more infrastructure should exist where privacy and sovereignty are foundational and people are in control. That quote is The Block's, from X. It is not a line in the Foundation post that loaded here.

The Block also says the GitHub repository describes the protocol as experimental. The Foundation page links the repository and does not use the word experimental in the text that loaded. If you need the warning label, it is The Block's reading of the repo, not a sentence on the blog.

On the design history, the pages split by a day. The Foundation says zkAPI is the working implementation of "ZK API usage credits," a design published on Ethereum Research by Davide Crapis and Vitalik Buterin. It does not print a calendar day in that background section. The Block says the same two people published the design on the Ethereum Research forum on Feb. 11, and its dek says "in February." Keep Feb. 11 as The Block's date. Do not paste it onto the Foundation sentence.

The Block also says the system does not provide network anonymity. It says the gateway can correlate requests from a stable IP, and that the post points at Tor for people who want a stronger network layer. The Foundation post does say network anonymity and content privacy are separate layers, and that a VPN or Tor handles the network side, with confidential GPU computing mentioned as emerging for content. So the Tor line for the zkAPI mainnet is in the post, and The Block repeats it. Sessions can still be re-linked by the content of the prompts, including personal details, writing style, or conversation history. The Block places that re-link risk with the post. The experimental label is a separate sentence about the repo. The payment layer not knowing your name does not make the prompt anonymous to the model provider. The provider still sees the request. That sentence is the whole design.

What the pages do not show

Neither page shows a fee table, a request count, or a list of model vendors that have turned the door on. The post explains how a provider would accept a proof. It does not say OpenAI, or any other lab, is a launch partner. Localhost compatibility means your app can speak a familiar API to software on your machine. It does not mean a hosted lab has agreed to forget you.

The vault "holding USDC credits" is a description on the Foundation page, not a dollar total. This article did not open the chain and count the vault. A reader who wants a zkAPI mainnet number should open the contract the post links, and should treat that balance as credits sitting in a vault, not as a market cap.

People tracking other AI shipping notes can read the pplx embed preview for a retrieval model, or the OpenAI Dots note for an agent launch. The zkAPI mainnet is a billing layer under tools like those, not a new model and not a scoreboard.

What to do as a reader (not a trade)

This is not investment advice and not a suggestion to deposit ETH or USDC into the vault. If you are evaluating the idea, separate three things the pages already separate. The payment layer sees spend and not your name. The provider sees the request. The network can still see your IP unless you add something like Tor, which the post names as a different layer. Read the experimental label as The Block's description of the repo, and read Feb. 11 as The Block's date for the research post.

If you cite the zkAPI mainnet, cite October 1, 2026, mainnet, Open Anonymity with the Foundation, nullifiers, and localhost OpenAI and Ollama compatibility. If you cite Liu, cite The Block. If you cite a dollar figure for the vault, you did not get one from these two pages.

The pages are the Foundation's post and The Block's story. Where the calendar differs, keep both. Where the provider still sees the prompt, do not let the word "private" erase that.