Credits
What a credit is, how a run is priced, what the platform fee covers, and how automatic top-up is bounded.
WARNING
There are no paid plans or credit purchases today. This describes functionality that is built but not launched.
A credit is prepaid access to model inference run through the NALA gateway.
Credits are not money. They are not a stored-value instrument, not a deposit, and not redeemable for cash. They cannot be transferred between accounts.
How a run is priced
A run costs what the model costs, at the rate published for that model.
credits = input tokens × input rate
+ output tokens × output rate
Two things about this worth knowing:
- A run is priced at the rate in effect when it started. If a model's price changes mid-run, your run is not repriced.
- There is no markup on inference. We do not add a margin to the token rate. Our revenue is the platform fee on purchases and unused allowance, both of which are visible to you.
You can see an estimate before a run and the actual cost after it. The estimate is deliberately generous, because reserving too little produces an overdraft, which is worse for you than briefly holding too much.
The platform fee
A percentage of a purchase, shown separately at checkout before you pay. The remainder is credited to your account.
It is charged on purchases rather than folded into the token rate so you can check the credits you received against the model's published price and see that the arithmetic is honest. A markup buried in a rate would be invisible.
Purchased credits do not expire
Unlike the weekly allowance, credits you buy stay until you spend them.
They are used only after the allowance is exhausted, so buying credits does not mean spending them while allowance remains.
Automatic top-up
Off unless you turn it on. If you do turn it on:
- You must set a spending ceiling. There is no "unlimited" option. A system that spends your money automatically without an upper bound is indefensible, so the ceiling is required by the code, not just the form.
- A top-up is trimmed rather than exceeding the ceiling. If topping up the full amount would go over, we charge the smaller remaining amount and tell you it was partial.
- Every charge notifies you. Money moving silently is the thing people cannot forgive.
- Repeated failures disable it rather than retrying a dead card forever.
You can turn it off at any time, with immediate effect.
When a run fails
If a request fails before the model generates anything, you are not charged and the reserved credits are released.
If a stream fails partway through, you are charged for the tokens that were actually generated. The provider bills us for those whether or not they reached you, and charging you the full estimate would be worse.
Where we cannot tell, we do not charge, and reconciliation against the provider's own billing corrects the difference the next day. That correction only ever runs in your favour: we refund an over-charge automatically and never retroactively bill you for an under-charge.