Agents and payments

Caching and refunds

Repeats cost half. Unused budget comes back.

Response cache

A byte-identical request can be answered from cache without touching a model. A hit is billed at 50% of the normal token charge and returns X-ArcRouter-Cache: HIT.

Deterministic by default. Only requests with temperature: 0 are cached automatically.
Opt in or out per request with the X-ArcRouter-Cache header: on caches at any temperature, off bypasses.
Tenant scoped. Your cache is keyed to your key or paying wallet. A hit never crosses tenants.
Non-streaming only.
Fails open. Any cache problem just sends the request upstream.

curl

curl https://arcrouter.co/v1/chat/completions \
  -H "Authorization: Bearer $ARCROUTER_KEY" \
  -H "X-ArcRouter-Cache: on" \
  -H "content-type: application/json" \
  -d '{"model":"gpt-oss-20b","messages":[{"role":"user","content":"Define x402 in one line."}]}'

Refunds

USDC is owed back to a wallet in four cases:

Overpayment. A pay-per-call ceiling minus the real charge.
Failed call. You paid and the model failed. The whole amount returns.
Failed deposit. A deposit settled but could not be credited.
Key closed. The remaining balance of a revoked key.

Refunds are batched per wallet and sent from the treasury on a sweep that runs every 10 minutes. Very small amounts wait until they add up to the minimum.

Empty replies are free

If a model returns no usable text, the charge is waived. On pay-per-call the whole ceiling is refunded.

NextErrors →