There is one number that turns up in almost every conversation about artificial intelligence: the price per token. It is falling, and has been for years. Anyone calling a language model through an API now pays a fraction of what the same amount of text cost not long ago. A simple conclusion follows: intelligence is getting cheap. Soon it will cost almost nothing.
The conclusion is convenient. It is also wrong – or at least incomplete.
Three prices that are not the same
When people talk about the cost of AI, they usually mean one of three very different prices and confuse it with the other two.
First, there is the price per token: the list price a provider charges for input and output. It is visible, comparable, and therefore what gets written about.
Second, there is the price per completed task: what does it actually cost to review a contract, find a bug, or answer a customer so that nobody has to redo the work? That price depends on how many attempts are needed, how long the context grows, how often a human steps in – and which model is good enough in the first place.
Third, there is the provider's full cost: data centres, chips, energy, cooling, staff, cost of capital. How much of that sits in the list price, and how much is carried by investor money, cross-subsidies or strategic losses, is hard to see from outside.
A low token price tells you nothing yet about the cost of a finished task.
When cheaper gets more expensive
The effect is paradoxical but well known. When a unit gets cheaper, the number of units consumed often rises. Systems that once produced one answer now produce ten intermediate steps, check themselves, call tools and write long justifications before they respond. Each operation is cheap. The task as a whole is not necessarily so.
And the cheapest model available is rarely the one used in practice. Whoever needs reliable results reaches for the more capable model – which seldom sits at the bottom of the price list.
Who carries the difference?
The real question, then, is not whether AI is getting cheaper, but who pays the gap between list price and full cost today – and who will pay it tomorrow.
As long as capital is cheap and growth matters more than margin, a provider can set prices that have little to do with its own costs. That is not an accusation; it is a familiar platform strategy: first reach, then retention, then pricing power. For companies building entire workflows on today's prices, it is a bet on the future that is rarely stated out loud.
The counter-argument
To be fair: efficiency gains are real. Smaller models keep improving, specialised hardware lowers the cost per operation, and competition limits how far prices can be raised. Those who can use several providers and keep models interchangeable genuinely benefit.
But that is exactly the point. The relief mostly reaches those who can switch. Those who cannot end up paying not the price of the token – but the price of dependence.
Note on method: this is a pilot essay. Concrete price comparisons will be published with date, task, quality requirement and model version.
