Designing the honest AI meter

Why Aardvark bills AI in dollars instead of credits, and the guardrails that keep a metered feature from ever producing a surprise bill.

Most products meter AI in credits — a currency you can’t reason about, converted at a rate you can’t see. When we designed Aardvark’s AI billing, we set one constraint up front: every number a customer sees is a dollar. Here’s what that constraint produced.

Published prices, real arithmetic. The pricing table lists what every model costs per answer on every plan — real figures, no conversion math. Paid plans include a monthly allowance denominated in those same billed dollars, and the dashboard shows dollars included, dollars used, your own trailing burn rate, and an estimated depletion date. No “≈N answers” marketing math.

The worst case is written down. The default when your included AI runs out is cap-and-hold: paid AI pauses, readers fall back to free models, and your max possible bill is your subscription price plus any overflow already incurred this month. If you opt into pay-as-you-go overflow, the dashboard also includes whatever remains under the cap you set — and raising that cap past 2× your allowance asks for an explicit acknowledgment, once, because a large overage should be a decision rather than a surprise. A third setting, shutoff, stops all AI including free models until the next month’s allowance arrives, for anyone who would rather the panel go quiet than keep answering.

Guardrails assume something will go wrong. An automatic safety brake pauses paid AI when spending runs far above your account’s normal pattern — one click resumes it, and it guards every account, free ones included, because a scraped embed or a looping script doesn’t check which plan you are on first. Concurrent paid requests are capped per account; past the cap a request gets a brief 429 with a Retry-After rather than queueing up a bill. And we’re precise about our own precision: the cutoff stops within roughly one request’s cost of your limit, not to the exact cent, and the docs say so.

Some smaller choices fell out of the same principle. Unused allowance rolls over one month, capped at about two months banked, so a quiet period can’t build into a bill-shaped balloon. Answers served by :free models never bill at all. And a paid plan’s metered AI never touches your prepaid balance unless you opt into overflow — on Free, where AI is pay-as-you-go, each metered answer draws that balance directly.

None of this is complicated. It’s mostly the discipline of refusing a layer of abstraction that only ever benefits the seller. The full details — every edge case, in plain language — live in the support knowledge base.