Most organisations now pay for AI twice: per seat for the assistants their people use, and per token for the systems their developers build. Nobody compares the two, so nobody knows which half is the expensive one.
Most organisations now buy AI in two ways at once. There are seat licences: a fixed monthly fee per person for an assistant like Copilot, ChatGPT, Claude or Gemini. And there is pay-per-use: the token bill for the systems the development team has built on the same vendors' models.
The two are bought by different people, from different budgets, on different terms. Finance sees one line for each. What nobody sees is whether the seats are worth more than the tokens they stand in for, and that is the question the next renewal should be answering.
Why can't you compare a seat with a token bill?
Because they price different things. A seat is a flat fee whatever the person does with it. A token bill moves with exactly what was asked and how much came back. The only fair comparison is to price what each seat holder actually does as if it had been bought by the token, from the same vendor, and see which number is lower.
Hold the vendor constant. Comparing one company's seat with another company's cheapest model is not a licensing decision, it is a migration, and it carries all the risk of one.
The break-even calculation
For any seat plan, the break-even point is the volume of use at which paying by the token would cost the same as the seat. Below it, the seat costs more than the work it carries. Above it, the seat is the cheaper way to buy that work.
You need four things for each group of users:
- How many requests a typical user sends in a month.
- How long those requests are, including any documents or files attached.
- How long the replies are, because output is priced at several times the rate of input across the major vendors.
- Which model the seat actually uses for that work.
Multiply it out at the vendor's own pay-per-use rates and you have the token-equivalent cost of each seat. The answer is usually not one number but a spread: a minority of heavy users who are well above break-even, and a long tail who are far below it.
What a worked example looks like
Take a team of 100 seats on a plan that costs the same per person every month. Suppose the usage data shows 20 heavy users whose token-equivalent cost is twice the seat price, 30 moderate users at about the seat price, and 50 light users at a fifth of it.
The heavy users are good value on seats: buying their work by the token would double what you pay for them. The moderate users are break-even. The light users are where the money goes: half the licences are carrying work that would cost 80% less on a pay-per-use plan.
Moving those 50 people to a pay-per-use arrangement, or to a cheaper seat tier where one exists, cuts the cost of that half of the estate by up to 80%, which is up to 40% off the whole bill for the team. The heavy users stay where they are.
Those numbers are assumptions chosen to show the shape. Your own split will be different, and the only way to know it is to measure it.
What a seat buys beyond the model
This is where most comparisons go wrong in the other direction. A seat is not only access to a model. It is also an interface people already know, single sign-on, administration, retention and data controls, support, and the vendor's commitments on how your data is used. A pay-per-use route has to provide all of that itself, and building it costs money.
So the honest output of the comparison is not an instruction to cancel seats. It is a verdict per user group, with the break-even point, and a note beside every verdict of what the seat provides that tokens would not.
Where the usage data comes from
Seat vendors provide admin reporting on activity per user. Pay-per-use vendors provide usage exports per key, project or workspace. Neither is designed to be compared with the other, which is why the comparison rarely gets done. Bringing the two into one view is most of the work.
Key takeaways
- Seats and tokens are bought separately and almost never compared
- The fair comparison prices each seat holder's actual work at the same vendor's pay-per-use rates
- Expect a spread: heavy users who justify the seat and a long tail who do not
- A seat also buys an interface, administration and data controls; say so beside every verdict
- The decision is per user group, not all or nothing
To run the arithmetic on your own numbers, use the free seat break-even calculator. It holds the vendor constant and splits your users into groups, as above.
This comparison is one of the four answers in the AI Spend Forecast, our three-week, fixed-price review of what your AI costs and what it should cost.