zkAPI Hides a Payment Link. It Does Not Hide Your Prompts
zkAPI separates funding deposits from metered API payments. The launch leaves prompt visibility, network correlation and provider integration as distinct questions.
Prepared automatically from the linked sources. The category byline does not confirm personal authorship or review. Editorial policy.
The Ethereum Foundation and Open Anonymity Project introduced zkAPI on October 1, 2026, offering a way to separate payment for metered APIs from the deposit funding it. The launch report identifies a system for AI models and other usage-based services. Its privacy claim concerns the payment link; it does not mean an inference provider cannot read a request.
Two paths put prompts in different places
The Foundation’s announcement distinguishes runtime-key and proxy modes. In the runtime-key flow, a payment server checks a spending proof and issues a short-lived, capped key. Requests then go directly to the inference provider. A signed usage receipt records the metered amount, which is deducted from the private balance when the key expires. The spending cap is a reservation, not necessarily the final charge.
In proxy mode, the zkAPI server relays requests to the provider and consequently sees traffic. The announcement separately warns that the provider sees request contents and network metadata. Reused personal details, writing style and conversation history can link sessions. A stable IP address and timing can also support correlation. Hiding which deposit paid for a session does not erase those other ways of connecting activity.
A research proposal and a working implementation differ
The February 11 Ethereum Research post, by Davide Crapis and Vitalik Buterin, identifies a challenge in combining privacy, security and efficient API metering. It proposes one initial deposit supporting repeated usage, with accounting for variable costs. The design aims to protect provider payment while keeping compliant usage unlinkable.
That research discusses refund tickets, a ticket index and a rate-limit nullifier to prevent reuse. It also explores an alternative method of accumulating refunds. These are descriptions of a proposed accounting design, rather than proof that every element appears unchanged in the later deployed system. Reading the February proposal as a current product manual would therefore miss the distinction between research and implementation.
The repository checked for this update documents a native ETH implementation. Its browser SDK covers funding, proof generation, settlement and withdrawals, with separate mainnet and Sepolia deployment pins. That current documentation is more specific than a general statement that a private-payment service exists on Ethereum.
The README also distinguishes local acceptance testing from external services: its lifecycle test uses mocked inference-provider usage and a mocked oracle, while exercising protocol services, wallet proofs and contracts. That is an explicit boundary of the described test. It does not establish that every commercial provider offers the required integration or independently demonstrate the behaviour of every deployed component. Server startup requires a pinned oracle RPC/feed and a lease issuer; the described acceptance test does not independently validate those external dependencies.
Compatibility is narrower than automatic availability
Editorial analysis: Payment privacy depends on the observer. A funding link, a relay’s access to content and a provider’s ability to connect sessions are distinct relationships. The documented limitations make that separation essential.
The Block’s launch coverage describes the implementation as experimental and live on mainnet. Those descriptions can coexist: deployment establishes availability of a system, while the reported experimental label identifies its development status. Neither description establishes universal anonymity.
The concrete missing information is deployment-specific: which mode and provider integration are operating, and what each observer can see. The billing design alone does not answer those questions.