DEV Community

Long AI
Long AI

Posted on

Why your LLM API invoice doesn't match the price page: a unit-by-unit guide

Every "your price page is wrong" ticket we open ends up being the same thing: two different units compared as if they were one. Text models are billed per million tokens. Image models per picture. Video models per second of output. Audio per minute of input. Four different rulers — and a single line that says "¥0.8" without the unit attached is a bug waiting to happen.

Here is the unit system behind the LONG AI price table, and the five places where a price page and an invoice can differ while both being correct.

1. Text is two prices, not one

Input tokens (system prompt, conversation history, tool definitions, retrieved context) and output tokens are metered separately, and output usually costs several times more. A page that shows one number is almost always quoting the input price.

Two consequences that surprise people:

  • An agent loop is not one call. Ten steps resend the history ten times, so input billing roughly multiplies by the step count.
  • Reasoning models bill the thinking tokens you never see. A short answer can carry a long invisible invoice.

When comparing providers, compare output price. That is where the spread between channels is widest, and it is the number that decides your monthly bill.

2. Images are priced per picture — and the tier is part of the price

The same image model often has several resolution or quality tiers at different unit prices. A quote taken at the cheapest tier sits next to a request made at 2K, and the invoice is arithmetically right. Pin the tier in the request, then compare prices at that tier.

3. Video is per second, audio is split in two directions

Video generation is billed per second of output, sometimes with extra charges for reference frames or upscaling. Five seconds at ¥0.3/s is ¥1.5 before extras. Speech synthesis goes per character or per second of audio; transcription goes per minute of input, so a long recording costs more regardless of how short the transcript is.

4. Currency is a settlement detail, not a hidden fee

A price quoted in CNY but settled in another currency moves with the exchange rate. That is why the English version of the site states the display currency and the settlement currency separately: displayed in USD, settled in CNY, with the rate visible.

5. Discounts are per model, not per site

"Up to 90% off" describes the deepest quote on the cheapest model in the catalogue. Your model has its own seller quotes, and those are the ones that apply. A site-wide average is useless for budgeting, which is why the model plaza shows per-model quotes with the channel behind each one.

Two smaller cases in the same family: retries (a failed upstream request should be billed once — if your client retries on its own, you made two billable requests) and tiers inside one model (long-context variants and peak multipliers change the unit price under the same model name).

The one-minute check

Before opening a ticket, write down four things: the unit (token / image / second / minute), the tier you requested, whether the counter is input-only or input plus output, and the settlement currency. If all four match between the price page and the invoice, the difference is almost always your own retry loop. If they do not match, it is a documentation bug worth reporting.

Every archived model on LONG AI lists its billing unit, its tier and both sides of the price for exactly this reason. Units are boring until they are expensive.

Top comments (1)

Collapse
 
aiapikey profile image
HAOHAOnovai •

Good list, and the per-minute-of-input point for transcription is the one that catches most people. Two additions, and the first comes from auditing a price feed rather than an invoice.

The third cause, beyond units and currency: the unit column itself can be a default value rather than a measurement. In the feed I audited this week, 48 of 116 rows carry the unit "flat", which is not a billing unit - and the distribution is diagnostic. flat is 71% of Chat, 73% of Audio, 67% of Embedding, and 0% of Image and 0% of 3D. It fires exactly in the categories whose real unit is per-token, and never in the categories where per-generation is genuinely correct, which is what a fallback looks like. Separately, 26 of 116 rows carry the identical price 1.54, covering every single product of four providers - one of them is 6 for 6. So before reconciling a page against an invoice, it is worth checking whether the page contains a price at all. A populated default looks exactly like data, and that is a third source of mismatch that unit-by-unit reading will not find.

The second addition is specific to your per-minute point. Billed audio minutes and input audio minutes are not always the same number. Providers that run voice-activity detection bill on detected speech rather than file duration, so a 40-minute recording with 12 minutes of speech can invoice as 12 or as 40 depending on the vendor and the mode. Neither is an error, and the difference is invisible unless the invoice line states which basis it used. That is worth a line in the checklist: does the invoice bill input duration or billed duration.

The display-versus-settlement currency split is the right way to handle it, and stating both with the rate visible is more than most relays do.