Reference

Limits and errors

What a call costs, what runs out, and what comes back when it does. The figures are read from the engine, which is the thing that enforces them.

What a call costs

Counted in answers; a reply drawing ten moments is one, fifty is five. A reply costs what it draws — the bars around each moment, its returns, where they sit against the usual — so the count follows the k or limit in your own query string. Ten moments is one answer and fifty is five, which means you can work out what a call costs before you make it.

The open endpoints draw nothing and cost nothing. Which is which is marked on every operation.

What each plan carries

freeFree
1,000 answers a month · 1 answer a second · 1 key
researchResearch
100,000 answers a month · 4 answers a second · $0.60 per 1,000 over that · 1 key
productionProduction
600,000 answers a month · 10 answers a second · $0.45 per 1,000 over that · 5 keys
redistributeRedistribute
600,000 answers a month · 10 answers a second · 5 keys

What a plan costs is on pricing; what separates them is what you may do with an answer, not only how many you take.

When something runs out

429. The body says which of the two it was — the month's allowance or the rate — and what had been spent, rather than only that something was too much. Where waiting is the answer, Retry-After is beside it.

A refused call is recorded too. A month of 429s is the clearest sign a plan is the wrong size, and it is invisible if only the answers that were given are written down.

The other refusals

400 the question could not be read. 401 no key, or not ours. 403 a real key asking for something that is not its own. 404 nothing of that name. 422 a parameter outside what it will take — the body names which. 503 the engine is up but has nothing honest to serve; /health says why.