Pay for the contents,
not the container.
On a scale, tare is the weight of the container — you zero it out so the price is for what is actually inside. We do that to a compute bill. We weigh the whole call, subtract what was never your work, and hand you the arithmetic.
8 models · 6 kinds of tare · no hardware attestation, and the receipt says so
Every deduction, with its reason.
One retried prompt and one stream the caller closed. Neither produced anything anybody asked for, and both are billed as usage by every gateway that meters gross.
That is the whole business: not a discount, not a credit — a description of what the number was actually made of. Add the middle column up yourself; it has to come back to the top one.
a real receipt, computed by the same engine the API runs — not a picture of one. conserved: true.
$0.009900 of $0.0288 was container
Everyone in this category meters the gross and calls it usage. The container is quietly in the price, and the only way you would ever find out is by adding a column nobody prints. We print it.
A bill you can re-add.
The container is not the contents
A retry after a 502 is our failure, not your work. So are the tokens a model burned before it errored, the stream you closed, and the seconds a sandbox spent booting. Everyone meters those. We weigh them, name them, and take them off.
The subtraction is checkable
Every receipt prints gross, each deduction with its reason, and net. net + tare = gross exactly, in whole micro-cents, and the identity is asserted on every response. A bill you cannot re-add is an invoice with a number on it.
Hard caps, enforced before the call
A budget cap per key, checked before work starts rather than discovered in a monthly total. A runaway agent stays cheap, and the receipt states what the cap did — including when it did nothing.
We do not sell a proof we do not have
We run on ordinary hardware. There is no confidential tier, no enclave, and no attestation claim, because we cannot prove one — and a tier that cannot be proven is a tier that eventually gets caught unproven.
Working today
- · the meter — 6 kinds of tare, each with a stated reason
- · exact conservation in whole micro-cents, asserted on every response
- · budget caps priced before the call, not after
- · the price sheet, public and unauthenticated
- · every receipt reproducible from its own numbers
Not live
- · the gateway that actually forwards inference
- · sandboxes you can run code in
- · API keys, prepaid balances, settlement
- · a confidential tier — and there is no plan to claim one we cannot prove
routing real inference needs upstream contracts and a balance ledger we have not shipped, so it is published as roadmap rather than dressed up as live. the part that decides what you owe already works, and you can run it yourself above on any numbers you like.