How word counts work on antigpt (and why failed requests are free)

Billing should be something you can predict and check. Here's exactly how we count, so there are no surprises.

What counts as a word

When you press Humanize, we count the words in your input using one rule, everywhere: normalise whitespace, split on it, and count each token that contains at least one letter or digit. So 'The results improved ... all 3 of them.' is 7 words: the ellipsis on its own isn't a word, and neither is any other standalone punctuation. The live counter in the editor uses the same rule as the server, so what you see is what you're charged.

We count input, not output. A rewrite that comes back longer than what you pasted costs you nothing extra; you paid for the words you submitted.

Reserve, then commit or release

Under the hood, submitting reserves the words against your balance. If the rewrite completes, the reservation is committed. If it fails (an upstream timeout, an error), the reservation is released and your balance is exactly what it was. This is why the interface can promise, honestly, that failed requests are free: it's built into how the meter works, not a goodwill gesture.

You can see it on the Usage & Billing page: released rows show as 'refunded'. Nothing is hidden.

Per-request cap vs monthly allowance

Two limits apply. Your plan's monthly allowance is the total words you can humanize in a period. The per-request cap is the most you can submit in one go: 300 on Free, up to 6,000 on Max. Over the cap, we offer to humanize the first N words or, on Pro and Max, to split the text at paragraph boundaries and stitch it back. Unused monthly words don't roll over, which we state on the pricing page.