api.prismnetwork.tech

One LLM generation. The price is the model's base plus its per-token rate over the requested output cap, and the unpaid 402 quotes the exact figure for every network it accepts. On Base, sign an EIP-3009 authorization for the quoted USDC amount and send it in the payment header; you need no gas, because the authorization is broadcast for you, and nothing is taken unless a generation is returned. A payment is consumed only when a response is served, and a consumed one replays its own result.

Many independent prompts in one paid call. The price is the single-request price times the number of prompts, and the unpaid 402 quotes the exact figure. Each prompt runs whole on one rented GPU, and the gateway spreads them over every GPU it holds, so a batch finishes in roughly the time the slowest box needs rather than the sum of all of them. The response carries a Merkle receipt over the set: each item comes with the digests it was committed under and an audit path, so any one answer can be proved to belong to the batch without disclosing the others, and the receipt names the leases that did the work, which settle on-chain with their own public receipts.

One chat completion served inside a Phala GPU TEE, in the OpenAI request and response shape. The price is the model's base plus its per-token rate over the max_tokens you ask for, and the unpaid 402 quotes the exact figure. The gateway forwards your request bytes unchanged and returns the upstream bytes unchanged, so the receipt the enclave signs over both covers exactly what you sent and exactly what you received; the id to fetch it comes back in X-Receipt-Id. Send the five X-E2EE-* headers to encrypt message content to the enclave's key, in which case the relay carries ciphertext. Set stream to true for server-sent events: frames are relayed as the enclave produces them, the receipt covers the whole stream including its framing, and the payment settles after the final frame, so a streamed answer carries no PAYMENT-RESPONSE header. A payment is consumed only when a complete response is served.

The TEE's attestation report for a caller-chosen nonce of 64 lowercase hex characters: the TDX quote, the measured compose, and the workload keyset the receipt signing keys and encryption keys are drawn from. Free.

The signed receipt for one completion, by the id its response carried. It binds the hash of the request bytes the enclave received, the hash of the response bytes it returned, and the upstream GPU verification outcome. Fetch it promptly; upstream retention is short. Free.

The attested upstream sessions currently serving, with the evidence each receipt cites. Add {id} to the path for one session's full record. Free.

The GPU attestation evidence for a model, in the shape NVIDIA's remote attestation service takes as a request body, so a caller can have the GPU leg verified by NVIDIA directly rather than by us. Free.

Models, per-model pricing, and the gateway's current state. Free.