Inference, priced per call.

An OpenAI-compatible gateway that charges per completion in USDG. Same request body as any Chat Completions endpoint, one difference: the first call comes back 402 with a price instead of 401 with a key error.

It is also the shortest way to watch a real payment settle. Fund a wallet, make one call, and read the transaction hash off the response header.

npx -y aeron-wallet address     # fund with USDG
npx -y aeron-wallet pay https://inference.aeron.sh/v1/chat/completions \
  '{"model":"deepseek/deepseek-chat","messages":[{"role":"user","content":"hi"}]}'

Endpoints

Prices

One flat price per call, not per token — the exact scheme needs a fixed amount before the work happens, and nobody can know the token count in advance. The flat price is derived from the model’s per-token cost at a typical call (2,000 prompt, 1,000 completion tokens), plus margin, doubled for variance, floored at $0.001 and rounded up to a tenth of a cent.

The catalog refreshes from upstream on a cron, so GET /v1/models is the only price list worth trusting. A sample of what it returns today:

curl -s https://inference.aeron.sh/v1/models | jq '.data[] | {id, flat: .pricing.flat_call_usd}'

{"id":"deepseek/deepseek-v4-flash","flat":0.001}
{"id":"openai/gpt-5-mini","flat":0.006}
{"id":"anthropic/claude-haiku-4.5","flat":0.015}
{"id":"google/gemini-3.1-pro-preview","flat":0.034}

Long calls can cost more upstream than the flat price collects. That is the gateway’s problem, not the caller’s: the price you were quoted in the 402 is the price you pay.

Paying without the wallet

The 402 body carries the offer. Sign it and retry — see signing it yourself in the wallet docs, or hand the whole flow to aeron-wallet.

curl -s -X POST https://inference.aeron.sh/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"deepseek/deepseek-chat","messages":[{"role":"user","content":"hi"}]}'

{
  "x402Version": 1,
  "error": "X-PAYMENT header is required",
  "accepts": [{ "scheme": "exact", "network": "eip155:4663",
                "maxAmountRequired": "4000", "payTo": "0x1b8e…80bd",
                "asset": "0x5fc5…d168", "extra": {"name":"USDG","decimals":6} }]
}

A successful paid call returns the completion plus X-PAYMENT-RESPONSE: base64 JSON holding the transaction hash. That header is your receipt — the same one Aeronscan indexes.

When it goes wrong

• Upstream failed before settlement — 502, and you are not charged. The authorization stays unspent.
• Upstream delivered nothing (empty generation, provider error) — the call is recorded as unpaid and you are not charged.
• Settlement failed — 402 with the reason. The completion is discarded rather than delivered unpaid.
• Payment invalid — 402 with the facilitator’s refusal reason and a fresh offer.

Streaming works: pass "stream": true and the response is passed through untouched while usage is metered from the tail.