Geeknewz exclusive explainer of Ethereum's new unlinkable API payment path, built from the Ethereum Foundation blog post (Oct 1, 2026), the February Ethereum Research design by Davide Crapis and Vitalik Buterin, zkAPI protocol docs, and same-day reporting from The Block. No invented partners, volumes, or quotes.
Most AI API bills still work like a credit card receipt stapled to every prompt. Your key points at an account, the account points at a payment method, and the provider can stitch years of sessions into one profile. On October 1, 2026, the Ethereum Foundation and the Open Anonymity Project put zkAPI on Ethereum mainnet to split that stack: the model provider still sees your prompts, the payment layer sees that a funded spend is valid, and neither side is supposed to learn the link between them.
This is not a new L1 fork or a privacy coin. It is prepaid credits plus zero-knowledge proofs for metered APIs, with AI chat listed as the first use case. Below is the short path from the February research note to a live vault, a plain map of who learns what, and a Geeknewz read on who should try it versus who should wait.
From research post to mainnet
The EF blog is explicit that zkAPI implements "ZK API Usage Credits," the design Crapis and Buterin published on Ethereum Research on February 11, 2026. Count the boring way: February 11 to October 1 is 232 days, a little under eight months from forum write-up to a mainnet vault and client you can point at localhost.
| When | What landed | Why it matters |
|---|---|---|
| Sep 2025 | EF forms the dAI team (per The Block / EF framing) | Ethereum as settlement for AI agents, not only DeFi |
| Feb 11, 2026 | Crapis + Buterin post "ZK API Usage Credits" on ethresear.ch | The design zkAPI says it implements |
| Jan 2026 | ERC-8004 agent identity standard on mainnet (EF / The Block) | Sibling rail for agent identity next to private payments |
| Oct 1, 2026 | zkAPI live on Ethereum mainnet with OA Project | Vault, client, OA Chat demo, Sepolia test deploy |
| Oct 6, 2026 (planned) | Glamsterdam on Sepolia (reported alongside zkAPI coverage) | Separate L1 upgrade track; not required to try zkAPI today |
That timeline is the Geeknewz connect-the-dots piece. Private AI payments sat in Vitalik's near-term "Ethereum for AI" talk earlier this year. ERC-8004 gave agents an identity primitive. zkAPI is the billing primitive that tries not to turn every inference session into a permanent customer dossier.
Who learns what (and who does not)
The EF post's own table is the clearest value-add, so we keep the structure and put it in plain English:
| Party | Learns | Never learns (by design) |
|---|---|---|
| zkAPI server | A valid payment exists; dollars spent per session | Who you are; what you asked; which deposit paid |
| AI provider | Prompts and responses (it runs the model) | Who is paying / which note funded the key |
| Ethereum public chain | Deposits, closes, withdrawals | What any private balance paid for |
Mechanically, you deposit ETH or USDC into a vault once. Your balance becomes a private note. To spend, software on your device builds a zero-knowledge proof that a funded note covers the spend and has not been double-spent. The server checks the proof, mints a short-lived, dollar-capped API key in your device memory, and your app talks to the provider with that key. When the key expires, metered usage comes off the private balance. Double-spends collide on a nullifier; honest spends stay unlinkable at the payment layer.
The client exposes OpenAI- and Ollama-compatible APIs on localhost, so existing editors and chat apps can point at it without a custom SDK. Funds sit in an onchain vault you can close and withdraw from even if every zkAPI server vanishes. That is the sovereignty claim, and it is checkable in the contracts the EF post links.
What zkAPI does not fix
Read the limitations before you treat this as Tor for ChatGPT. The EF post and protocol docs are blunt: zkAPI is not network anonymity. A stable IP at the gateway can still correlate sessions. Tor (or another network layer) is a separate choice. Content can re-link you too. The same personal details, writing style, or pasted project files can fingerprint sessions at the model provider even when the bill is unlinkable. The GitHub and docs also call the system experimental. The current elliptic-curve setup is not post-quantum, and the checked-in trusted setup is single-party.
Use cases beyond AI chat are listed the same way: RPC queries, image and video jobs, VPN bandwidth, and machine-to-machine agent payments. Same client pattern, different metered unit. Pricing and rate limits stay with the provider; integration means accepting proofs and settling signed usage receipts instead of long-lived account keys.
Geeknewz verdict
Geeknewz's view: if you build agents, research tools, or anything where the billing identity is the privacy leak, zkAPI is worth a Sepolia or small mainnet deposit test this week. The under-eight-month path from ethresear.ch to a live vault is unusually concrete for ZK payment designs. If you only want cheaper tokens or a better model, this changes nothing about quality or price. And if you need "nobody can tell I used the API," you still need network anonymity and careful prompt hygiene on top. Treat unlinkable prepaid credits as one layer, not a full privacy suit. Watch whether major providers accept the runtime-key path without a relay that sees traffic, and whether the experimental label graduates into something your security team will approve.