The situation
You paid for a twenty-dollar plan on Monday. On Thursday afternoon, halfway through the one job that mattered this week, it stops: limit reached, try again in a few hours. So you read that you can supply your own key instead and pay only for what you use, which sounds cheaper and more honest. It is, right up until the Saturday that costs more than the month you were trying to avoid.
The idea in one paragraph
There are two ways to buy the same thing, and they fail in opposite directions. A subscription buys a share of capacity: a flat price, and a meter you cannot see that stops you once you have had your share. Paying by usage buys the work outright, priced by the volume of text going in and coming back: nothing stops you, and the bill arrives afterwards. Neither is the cheap option. Which one hurts depends on how heavy your worst day is and how long you let one session run.
Where the money actually goes
| How you pay | What you get | What stops you | The surprise |
|---|---|---|---|
| Nothing | A free tier | A hard cap, reached fast | Enough to evaluate, not to work |
| Flat monthly | A share of capacity | A limit that resets on a clock | You cannot watch the meter |
| A heavier flat tier | A larger share | The same limit, later in the week | A multiple of a number nobody shows you |
| By usage, on your own key | Text in, text out | Nothing at all | Every turn re-bills the whole conversation |
Only the last row can charge you for a mistake while you are still making it.
How it actually works
Prices checked on 10 August 2026, in dollars before tax. Entry subscriptions for one person sit between $10 and $20 a month: GitHub Copilot Pro at $10, Cursor Pro at $20, Claude Pro at $20 or $17 a month if you pay $200 for the year. The heavier tiers run $39 to $100: Copilot Pro+ at $39, Copilot Max and Claude Max both starting at $100. All three have a free tier.
Usage pricing is quoted per million tokens, a token being the chunk of text a model reads or writes, around four characters of English. One vendor's current models run $1 to $10 per million tokens in and $5 to $50 per million out. Output costs roughly five times input across that whole list, so you pay the high rate on the long answer rather than on your short instruction.
The part that catches people out: a conversation holds no memory of itself, so every turn re-sends everything before it. Turn ten pays for turns one to nine again. Vendors discount a repeated opening (a cached prefix is re-read at a tenth of the input price) but that discount is designed for a prefix that stays put and it expires in minutes. A long session is not linear: it is the same growing pile of text, bought again on every message.
What this changes for you
- Start on a flat plan, not a key. When you get something wrong there, it costs you a wait rather than a number, and a wait teaches more than a month-end statement does.
- Keep sessions short and finishable for the money as well as the accuracy. What sits in the context window is re-read and re-billed every single turn.
- If you do use a key, set the spend limit in the account before the first call, not after the first surprise. A tool that will not let you set one has told you something.
Where it breaks
Every figure above is a dated snapshot. Plans get renamed, limits get retuned, and introductory rates end. Check the vendor's own page before deciding, and treat any number you read on this subject, this one included, as expired until you have.
Cheapest per token is not cheapest per finished job. A weaker model that needs four attempts and a long correcting conversation can cost more than one pass from an expensive one, and it also costs your afternoon.
The meter is the small cost. The large one is the hour you spend reading, testing and undoing what came back, which no plan prices and no free tier removes.
Terms used on this page
- Usage limit — the cap a subscription enforces inside a rolling window. It stops you and then resets; it does not bill you.
- Usage pricing — paying by the volume of text instead of a flat fee, quoted per million tokens. There is no ceiling unless you set one.
- Input and output tokens — text sent versus text produced, billed at different rates. Output costs several times more.
- Prompt caching — a discount on an opening the model has seen recently, re-read at a fraction of the normal input price for a short window.
- Spend limit — a ceiling you set on an account so a runaway loop stops rather than bills.
Read next
Chat versus agent — what changes when the model can read your files, write to them and run commands, and what you are agreeing to each time you allow it.