Subscription or pay-as-you-go: which costs you less?
Subscription or pay-as-you-go
Two ways of paying for the same thing. The maths runs in your browser.
Based on 220 working days. The price per million tokens deliberately merges input and output: an approximation, good enough to choose between the two models, not to budget to the cent.
The parameter that decides everything is neither the subscription price nor the token price. It is the share of people who genuinely use the tool. Equipping fifty people of whom twelve open it daily means paying for thirty-eight dormant seats.
The two models, and who they suit
| Per-person subscription | Pay-as-you-go | |
|---|---|---|
| Billing basis | A monthly rate per head | The volume of text processed |
| Predictability | Total | None without a cap |
| Heavy daily use | Cheaper | Expensive |
| Occasional use | Dormant seats | Cheaper |
| Automated processing | Not applicable | The only option |
What a token is, and why the count surprises
A token is roughly a fragment of a word. Three things explain why the bill almost always exceeds the estimate.
- What you send is billed as much as what you receive. The instruction, the attached document, all of it counts.
- The history is resent every turn. A long conversation costs more with each exchange, since the whole context goes back.
- Models do not cost the same. The gap between a fast model and a reasoning model is measured in orders of magnitude, for sometimes equivalent results on a simple task.
Billing traps
- Set a spending cap on the very first call. A badly written loop produces a bill unrelated to intended use.
- Remove dormant seats after three months. The simplest saving, and the most often neglected.
- Do not forget verification time. It appears on no invoice and often weighs more than the licence.
For rates product by product, see our 2026 AI comparison.
Frequently asked questions
What share of real users should I use?
Measure it rather than estimate it. On a broad rollout without support, the share of people opening the tool daily is usually far below what management imagines. That is precisely why this field sits at the top of the calculator: it weighs more than all the others combined.
Does this hold for automated use?
For an order of magnitude, yes, provided you translate your volume into requests and tokens. The per-person subscription no longer makes sense there, since no person is doing the consuming. Only the pay-as-you-go column matters, and a spending cap becomes mandatory.
Why merge input and output tokens?
Because the ratio between them depends entirely on your usage, and a calculator asking for both separately would give a false impression of precision. The approximation is good enough to choose between the two billing models, which is the question being asked here.