Private AI Inference, Paid Per Request

x402 is an open payment protocol built on the HTTP 402 status code. A server answers an unpaid request with 402 and a PAYMENT-REQUIRED header that describes the price, the asset and the network. The client signs a payment authorization, retries with a PAYMENT-SIGNATURE header, and a facilitator settles it on-chain. No accounts, no sessions, no card forms.

The gateway holds your prompt and the answer in memory only while the request runs. It writes no logs of content and stores nothing afterwards; the receipt carries a SHA-256 of the prompt and the token counts, never the text. Today this is a property of open-source code you can read. The roadmap moves the gateway into a hardware enclave so that it becomes a property you can verify from the attestation in every 402.

The signed receipt reports the exact model id that produced the answer, as returned by the provider, together with the token counts you were charged for. Once the gateway runs inside an attested enclave, the build measurement in the 402 will bind that promise to hardware, and a client can refuse to pay any other build.

With "exact", the price is fixed before the response. With "upto", you authorize a ceiling and the node settles only the tokens it produced. A 2,000-token ceiling that returns 188 tokens charges 188 tokens. Streaming works the same way; the receipt arrives after the last chunk.

Not by Umbra. The settlement record holds the prompt hash, not the prompt. For stronger unlinkability, use a fresh address per task, or Umbra Relay, which routes the request through a mixing relay so the node never sees your IP alongside your signature.

Umbra runs a single gateway today. It settles in USDG on Robinhood Chain (mainnet, real money, the gateway facilitates its own settlements), and in test USDC on Base Sepolia so anyone can try the full paid flow for free. Prices are Anthropic list prices plus a 4% protocol fee. Independent operators on confidential hardware come after the enclave build ships.

An authorization is not a payment. If the node fails to answer, nothing settles and the authorization expires in 60 seconds. If it answers partially under "upto", you pay for what you received, and the receipt shows exactly that.

Two things: the base URL, and the fetch implementation, which you wrap with an x402 signer bound to a wallet. Model names, streaming, tool calls and vision work exactly as they do today with the OpenAI and Anthropic SDKs.