You do not need a credit card to start calling GPT through a gateway. Prepaid access works the other way around from a vendor console: instead of attaching a card and being billed postpaid at the end of the month, you load a small balance up front and that balance is drawn down token by token. When it runs low, you top up; while it lasts, you owe nothing more.
That single difference — prepaid instead of postpaid — removes the card requirement and changes what a mistake costs you. A postpaid key can run up a bill you only see later. A prepaid balance can only ever spend down to zero, so the worst case is a number you chose yourself.
This is the main reason a prepaid gateway is the practical answer to the "no card" search. A vendor console wants a card up front because it will bill you after the fact; a gateway asks for nothing at signup because the risk is bounded by the balance you load. For a developer testing a side project or a small team that wants to cap spend, the prepaid shape is both simpler and safer than a card on file.
Prepaid is a running account balance measured in the same tokens you are billed for.
Nothing about the API itself changes because you are prepaid. You still use the same OpenAI-compatible endpoint, the same request shape, and the same model names.
Three policy rules are the reason prepaid is low-risk, and they are worth checking against any provider you consider.
| Rule | What it means |
|---|---|
| Credit does not expire | Unused balance stays on your account; there is no clock running it down |
| Failed requests are not charged | A request costs money only when tokens are actually generated |
| 14-day refund window | You can request a refund on unused balance within 14 days if you change your mind |
The first two matter in everyday use. The third is the one that matters if you try a provider and decide it is not for you.
Starting is three steps, and none of them asks for a card.
/v1/chat/completions request to confirm the key works.If that one call returns output, you have a working path to GPT models without a card in the loop. If it does not, you have spent nothing — the failed request is not billed.
Prepaid without a card is not the same product as a vendor's postpaid console, and two differences are worth naming.
For an independent developer or a small team that wants to control cost and skip the card, those are usually acceptable trade-offs. For a procurement-led organisation, they are not — and that group should buy direct.
One more thing worth knowing before you commit a real workload: prepaid access is drawn by usage across whatever models you call, so the same balance covers GPT and any other model on the same endpoint. You are not opening a separate prepaid account per model family — one balance, one key, and the per-model rate decides how fast it drains.
Q: Do I actually need a credit card to get a GPT API key?
Not through a prepaid gateway. You create an account and load a balance without a card requirement at signup, and the balance is what you spend.
Q: Does my unused credit expire?
No. Unused prepaid credit stays on your account and is not run down by any expiry clock.
Q: Are failed requests charged?
No. A request costs money only when tokens are generated, so a failed or retried request that produces no output is not billed.
Q: What if I load money and then decide not to use it?
There is a 14-day window to request a refund on unused balance. After that, the balance simply stays available until you use it.
Q: How much do I need to start?
The minimum top-up is $10. There is no monthly fee and no minimum spend beyond that initial load; you pay only for the tokens you use.
1. Topping up before testing. Prepaid only helps if you test first. Run a handful of requests on whatever free credit you have, read the usage block, and confirm the per-token arithmetic matches the price page. A $10 top-up before that step just moves the risk onto you.
2. Treating the balance as a subscription. Prepaid credit is not a monthly plan: nothing renews, nothing expires, and an idle month costs zero. The failure mode people worry about — "I paid and then stopped using it" — only matters with subscriptions, which is the thing you avoided.
3. Assuming failed requests are billed. They should not be. A request that errors before generating tokens has nothing to charge for. If your statement shows a charge for a failed call, that is worth a support ticket — and it is the kind of thing a provider should be able to explain line by line.
None of this is specific to us. It is the arithmetic of paying in advance instead of signing up for a plan, and it works the same at any provider that lets you check.