Getting started
Questions, answered
What the bill covers, what happens when the balance runs out, which library to use, and what we keep.
Pricing and fees
What exactly am I charged for?
Tokens, on four counters separately: input, output, cache write and cache read, each at the catalogue rate of the model the request used. Image models are metered in tokens too: the prompt and any attached images are input, the frame is output, and how many output tokens a frame takes depends on its shape. Nothing else is metered.
Are there any additional fees?
One: a service fee on top-ups — 5% of the credited amount, added on top and shown on the top-up screen before you pay (credit $50.00, send $52.50). Nothing else — no platform percentage on requests, no monthly minimum, no subscription.
Credits and billing
What happens when the balance runs out?
The next request gets a 402 payment_required and nothing is charged. Top up from your account and continue — nothing in your settings changes.
Do credits expire?
No. An idle balance has no shelf life and no inactivity fee.
Can I get a refund?
Yes. Unused balance is refunded on request, within up to 10 business days — write to support@gtokens.io from your account email.
Integration
Which client library do I use?
Either of two. An OpenAI-compatible client pointed at the /api/v1 base URL reaches every model: text models answer on chat/completions with tool calls and streaming (the GPT models answer with text only, as their pages say), image models return frames there and on the edit route. A Messages API client pointed at the gateway root, with the key as the bearer auth token, reaches the other text models in their native shape. The Quickstart shows both setups; one key serves both, only the base URL differs.
Do I need an account with the model vendor?
No. Requests are served under our own access; your key only needs to identify your account here.
Models
Why do some models have no price?
They are marked coming_soon in the catalogue: listed, not served. A request naming one comes back 400 model_not_available, and a refused request costs nothing. No model is ever silently replaced by another.
Requests and errors
Is there an uptime guarantee?
No — we promise no uptime figure, and the terms say so plainly. What we do publish is the state of each part on the status page, together with the record of every incident that reached customers. When our pool for a model is fully busy the request gets a 503, with a Retry-After header where the pool can name one: honour that interval, otherwise wait a few seconds and retry. A refused request costs nothing.
Account and keys
How many keys can I have?
As many as you need. They share one balance and any ceilings set on the account; each can be revoked on its own in the account area.
About
Who is behind this?
A private project, not a legal entity: no contracts, no invoices. If you need a paper trail, this will not suit you.
What we store
We do not store the contents of your requests or of the model’s replies — not for debugging and not for analytics. This is a property of the system rather than a promise of goodwill: bodies stream through the gateway and settle nowhere.
- the model, the four token counters or the number of frames, and the charge for each request;
- the latency, the timestamp, and which of your keys was used.
One exception, and it is about someone else’s text rather than yours: when the upstream provider refuses a request, the text of that refusal is kept against our own upstream account and in the server log, because otherwise the cause cannot be established afterwards. Your own request text goes into neither.
Your requests go to the upstream, because there is no other way to answer them; their data-handling rules apply on their side and we do not influence them. They go nowhere else, and the site itself carries no third-party analytics, visitor counters or advertising scripts.
The honest flip side: because contents are not stored, there is nothing to produce in a billing dispute except the counters. That is why every line of your log carries its four counters and its charge, and why the billing section of the documentation shows the arithmetic behind each line.