A token pack if you know you are about to do a lot of work in the next day or three: it is far cheaper per token, and the short validity window is exactly what pays for that. Credits if your usage is unpredictable or occasional — they cost a little more per token but sit on your balance indefinitely. Plenty of customers keep a credit balance as a floor and buy a pack when a sprint starts. When both are present we always spend the soonest-expiring pack first, so nothing is lost by accident.

Pay as you go, metered per token at the published per-model rate. Cost is computed at full precision and then rounded up to the nearest 0.000001 credit — one micro-credit. Streaming requests place a temporary hold for the maximum possible cost and settle to the real amount when the stream ends; a cancelled stream is billed only for the tokens actually delivered.

Yes. A pack is a flat bucket — input, output and reasoning tokens all draw from the same number, with no 5× multiplier on output. That is where most of a vendor bill normally goes, which is a large part of why a pack works out cheaper than metered credits.

One credit equals one US dollar of vendor list-price usage. You buy credits at 85% off list, so $45 buys 320credits. That keeps the price table directly comparable to the vendor's own published rates — the discount lives in what you pay for credits, not in a rewritten price sheet.

The clock starts when your payment confirms, not at first use, and it runs on wall-clock hours: a 2-day pack bought at 14:00 Monday expires 14:00 Wednesday. The exact timestamp is on your billing page. At expiry 10% of whatever is left converts to permanent credits, so an expired pack still leaves something on your balance. You can stack multiple packs — they add together, and we always spend the soonest-expiring one first.

Credits never expire. Token packs do — that is the trade-off for the lower price per token, softened by the 10% rollover described above. Unused credits can be refunded within 14 days of purchase, less any usage already metered; an expired pack is not refundable. Full terms are in the refund policy.

Requests per minute, tokens per minute, requests per day and concurrent streams are all enforced per key, and the current values are shown on your dashboard. There is no daily request cap on paid balances. When you hit a limit you get a 429 with an accurate Retry-After header — never a silent stall.

By default we store the full request and response payload, alongside your user id and IP address, for 30 days. That is what lets us debug a failed call with you. We do not claim zero data retention while that is on. If you would rather we kept only metadata — token counts, model, latency, cost — switch on Zero Data Retention in Settings; from that moment payload storage is skipped for your account.

No. FuelOfAI is an independent API aggregator and is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, or Google. All trademarks belong to their owners. We buy upstream capacity with our own paid accounts and resell metered access to it.

Each model has an ordered list of upstream channels. A timeout, 429 or 5xx moves the request to the next channel automatically, and a repeatedly failing channel is taken out of rotation by a circuit breaker until it recovers. Once a stream has started sending bytes we do not switch mid-flight — you get a terminal error event instead, and you are billed only for what was delivered.

Only models we have verified on a live channel appear in the catalogue and in GET /v1/models. We would rather publish a short honest list than a long aspirational one.