K2Coder can use AI in two ways. With your own keys, you pay the AI provider directly and none of this applies. With Quantum Cloud credit, you buy prepaid credit from us in US dollars and K2Coder draws on it as it works. This post explains exactly how it draws on it, because a builder that runs for minutes on its own should never leave you guessing what it spent.
Some of the rules below are new. They are built and tested, and they take effect with the next update of the Quantum Cloud service. Each one is marked.
Every request: set aside first, charged exactly after
Each time K2Coder asks a model for something, two things happen.
Before the request, the most it could cost is set aside. We work out the worst case: the size of what is being sent, plus the longest answer the model is allowed to write, at that model's price. That amount moves from your available credit into a hold. If there is not enough credit, the request stops there and the model is never called.
After the request, you are charged what it actually cost. The provider reports how many tokens it read and wrote; we charge those at the model's price, and the rest of the hold goes straight back to your balance. The charge can never be more than the hold.
Two more rules sit around this:
- Only models with a checked price can be paid for with credit. If we have not confirmed a model's price, it is not offered on credit.
- Claude models are never sold on credit. They use your own Claude plan, under Anthropic's terms.
A project can have its own budget. When a request would take a project past its budget, it is refused before anything is spent.
The model never writes more than was set aside (new)
The hold assumes a longest possible answer. From the next update, the request sent to the model carries that same limit, whichever provider it goes to. A model cannot write a longer answer than the one you were held for.
Every request is settled on the spot (new)
From the next update, a request that runs too long is stopped after about four and a half minutes and settled immediately, rather than left open, so its cost is recorded straight away.
If a request is cut off after the model already received it, because it ran out of time, the app closed, or the connection dropped, you are charged an estimate of what the model read and what came back before it stopped. The provider charges us for that work either way. If the request never reached the model at all, nothing is charged.
A request is never sent twice when it might already have run (new)
Sometimes a provider is busy and turns a request away. From the next update, K2Coder only tries again when the provider has said it did not take the request, or when the connection never opened. If there is any chance the model already did the work, it does not send the same request again. Either way, you are charged once per request.
When you are refunded, the credit goes back (new)
If you ask for a refund on a credit purchase and it is processed, the matching credit comes off your balance. For a partial refund, the matching share. The amount comes from the payment provider's own records, and each refund counts once.
If some of that credit was already spent on builds, the remainder is recorded and your account is paused until it is settled with us. The same applies while a card dispute is open. We would rather tell you this now than have it come as a surprise.
Where to see it
Settings → Account in K2Coder shows your balance, what each build cost and each project's budget. Questions about a charge: support@quantumaccordbg.com.