The six pricing shapes
Once you recognise which one you are dealing with, the rest is arithmetic. These six are in circulation today, and most providers combine two or three of them.
- Tiered subscription. A fixed monthly fee, with different usage limits and different models per tier. This is the base model of chat interfaces.
- Usage-based, per-token pricing. You pay for what you consume. This is the base model of APIs.
- Credit pools. You buy a balance up front, and different operations drain it at different rates. Common in developer tools.
- Outcome-based pricing. You pay not for usage but per case resolved. Already established in customer-support AI — in places it is under a dollar per closed case.
- Seats plus an AI surcharge. The existing software's per-user price, with an AI package laid on top.
- Freemium. A free tier with limits, a paid tier for lifting the limits.
Token pricing, where most of the money goes
API pricing has two numbers, and most people only look at the first: input and output tokens. What you send in is input; what comes back is output.
Output always costs more, typically four to five times more. That matters because intuition works the other way round: you would think uploading the long document is the expensive part, when in fact the long answer is.
From which follows the simplest saving anyone can make: ask for a shorter answer. Not because it hurts the model, but because output is the pricier half of the bill.
Four multipliers the list price does not show
The list price is a base case. Four things move it, and all four will be familiar:
- Reasoning. When the model thinks, the thinking counts as output tokens too — even when you never see it. The same question on a reasoning tier can cost several times as much.
- Language. Hungarian text breaks into 2.71 times as many tokens as English of the same length; Spanish 1.46 times. The same document therefore costs nearly three times as much in Hungarian.
- Caching. When the same prefix repeats, providers bill it at a fraction — typically a tenth. This is not automatic; you have to ask for it.
- Batch runs. If you accept that the answer arrives within hours rather than within seconds, in many places it costs half.
A snapshot from September 2026
These numbers will be stale in six months; the ratios will not. Per million tokens, in dollars:
| Tier | Input | Output |
|---|---|---|
| Top tier | 5–10 | 25–50 |
| Middle tier | 2 | 10–12 |
| Cheap tier | 0.2–1 | 1.2–5 |
| Open-weight, cheap | 0.14 | 0.28 |
What survives from this: the gap between the cheapest and the most expensive tier is twenty-five fold, and the cheap open-weight line sits orders of magnitude below even that. This is why picking the tier is worth more than any amount of haggling.
The hidden costs
The conversation history. In every turn, the whole conversation so far goes through again as input. A twenty-turn conversation does not cost twenty times as much, it costs far more — it grows quadratically.
Retries. If the answer is bad and you run it again, you paid twice. In an automated process this is built in, and nobody watches it.
The system instruction. If a two-thousand-word instruction and five examples ride along on every call, you pay for them on every single run. At ten thousand calls a day that line item can exceed the actual work.
The unused seat. On a subscription model the biggest waste is not usage but non-usage: the licence that was handed out and never opened.
How to estimate in advance
You do not need a spreadsheet, you need one real measurement. Four steps:
- Take a typical task — not the shortest and not the most extreme.
- Run it five times and look at the actual token consumption. Most providers report it per call.
- Multiply by the monthly volume. That is the raw base.
- Multiply by two. Retries, longer cases, growing use — estimates always come in under.
If that number is larger than the task is worth, the answer is not a more expensive model but fewer tokens: shorter context, a smaller tier, fewer runs. That is the practical payoff of context engineering.
What to watch in the contract
What happens on overage? Does it stop, throttle, or just bill? Those are three very different risks.
Can the price change mid-term? Prices move — there have been cases of a provider doubling an introductory rate later. If you are signing an annual commitment, that is the sentence that matters.
What counts against the allowance? Reasoning, tokens served from cache, failed calls — providers count these differently, and the difference is not small.
What to take away
Output costs more than input, and the tier matters more than the vendor. Four things move the list price: reasoning, language, caching and batching. And always start the estimate from a real measurement, then multiply by two — the surprise always comes from above.