There are two ways to buy the same thing. A subscription stops you when you have had your share, usually on the afternoon you needed it. Your own API key never stops you, and the bill arrives on Saturday.
The idea
A subscription buys a share of capacity: a flat price, and a meter you cannot see that cuts you off and resets on a clock. Paying by usage buys the work outright, priced by the volume of text going in and coming back. Nothing stops you there. Neither is the cheap option. Which one hurts depends on how heavy your worst day is and how long you let one session run.
| 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 clock, reached later | You are buying a multiple of an amount never shown to you |
| Your own A secret string that bills your work to your own account, with no monthly plan and no cut-off. You paste it into the tool once. | Text in, text out | Nothing at all | Every turn re-bills the whole conversation |
How it works
Prices checked on 13 August 2026, in US dollars before tax.
- Entry plans for one person run $10 to $20 a month. GitHub Copilot Pro is $10, Cursor Pro is $20, Claude Pro is $20 monthly or $17 if you pay $200 for the year.
- Heavier single-person tiers run $39 to $200: Copilot Pro+ at $39, Cursor Pro+ at $60, Copilot Max at $100, Claude Max at $100 or $200 depending on the share you want, and Cursor Ultra at $200.
- All three have a free tier.
Usage pricing is quoted per million The chunks text is split into before a model reads it. Anthropic puts one token at roughly four characters of English.. Anthropic's published API pricing runs $1 to $10 per million tokens in, and $5 to $50 per million out, depending on the model. Across Anthropic's current list output is priced at exactly five times input, so the high rate lands on the long answer rather than on your short instruction. Other vendors do not all use that ratio.
Here is the part that catches people. A conversation holds no memory of itself, so every turn re-sends everything before it. Turn ten pays for turns one to nine again. Anthropic discounts a repeated opening (a The opening stretch of a request the vendor recognises from a recent one and re-reads cheaply. Change anything inside it and the discount is gone. is re-read at a tenth of the input price). That discount lasts an hour on a Claude subscription, and five minutes on an API key or once you are drawing on usage credits.
What a session costs is dominated by re-reads. Anthropic prices 50,000 tokens in and 15,000 out on Opus 5 at $0.25 plus $0.38, about 63 cents, and that is one short exchange rather than an afternoon. A real session re-sends its conversation every turn, so a few hours runs to hundreds of thousands of re-read tokens, most billed at a tenth of the input price when the cache holds. Anthropic's own published figure for teams running Claude Code is around $13 per developer per active day.
What to do
- Start on a flat plan rather than an API key. A mistake there costs you a wait instead of a number, and a wait teaches faster than a month-end statement.
- Keep sessions short for the money as well as the accuracy. Whatever sits in the context window is re-read and re-billed on every turn.
- If you do use a key, set the spend limit in the account before the first call. A tool that will not let you set one has told you something.
Where it breaks
Every figure above is a dated snapshot, and this subject moves fast enough that one of these prices changed while the section was being checked. Confirm at claude.com/pricing, cursor.com/pricing and GitHub's plans page before you commit money.
Cheapest per token is rarely 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 costs your afternoon as well.
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.