Let’s talk about the bill that went up while the price of the thing on it went down.
For two years the story has been identical. Token prices falling. DeepSeek put a knife through the floor, then OpenAI, Google and Anthropic followed, and every launch deck since has had the same chart — a line going down and to the right, with a promise underneath it that intelligence is getting cheap.
It is. Per unit. Nobody’s arguing with that. The bill is the part nobody reads.
Picture a small studio. Three people, brand work for mid-sized companies. They quote a fixed price for a fixed deliverable, because that’s how the client’s finance department works and how the studio has always worked. Forty thousand rand for the campaign, six weeks.
They build an agent to do the boring middle. Pull the brief apart, draft variants, check the copy against the brand guide, resize the assets. It works. And on the first proper job the meter runs to nine thousand rand, because the agent read the same sixty-page brand guide eleven times, retried the image step four times, and rebuilt the entire plan every time the designer corrected a headline.
The client pays forty. The studio pays nine. Nobody lied, nobody overcharged, and the job made less money than the one they used to do by hand.
That’s the shape of a trade nobody in the room actually agreed to.
In August, Gartner put a name on it — the Inference Paradox. Better unit economics escalating the total cost of AI without a clear path to value that matches. Their call: inference cost per agentic workflow rises more than fivefold through 2028.
And the reason is mechanical, not mysterious. A chatbot reads a question and answers it. An agent reasons, negotiates with itself, calls a tool, checks its own work, and does the whole thing again when a step fails. Routing one task to a reasoning agent costs a provider at least five times what a chatbot interaction costs, and that multiple climbs as the task gets more complicated.
Their analyst line is the one to tape to a monitor. Defaulting to generic autonomous intelligence produces unbounded costs, orders of magnitude above an optimised system.
Five times isn’t a rounding error. It’s a different business.
This isn’t a new invention either. The copier contract did it in the eighties — cheap machine, a charge per page, and the dealer knew your monthly volume before you did. Telecom did it with minutes. Cloud did it with egress. Every one of those was sold as “you only pay for what you use,” and every one of them ended with a customer who couldn’t forecast a bill they had no instrument to measure.
AI just moved the meter from pages to thinking.
So look at what the pricing pages did while everybody was toasting cheap tokens. A census published this week counted 42 dated pricing changes across 23 agent vendors since August 2024. Five were plain increases. Seven were decreases. The other thirty — and this is the whole story — restructured what a plan includes or introduced a new unit.
Intercom renamed the unit it bills on without changing the number. Salesforce has priced Agentforce four different ways in under two years. GitHub moved Copilot off request allowances and onto AI credits and left every seat price exactly where it was.
That’s the quietest price rise in software. The sticker holds still and the unit moves under it. Customers compare stickers. The invoice arrives in the new unit.
Fifty-six percent of the agent products in that census now bill on a usage meter. Thirteen per credit, six per run or task, four per minute. Four per outcome. Four per seat.
Here’s what a meter actually is. It isn’t a pricing model. It’s a risk transfer.
On a seat, the vendor eats the cost of its own inefficiency. If the tool burns twice the compute it should, that comes out of their margin. On a meter, it comes out of yours. And a language model is inefficient in precisely the places you can’t see — the retry after malformed JSON, the context window re-read on every turn, the plan it abandons and rebuilds, the tool call that returns nothing useful three times before it gives up.
You are not paying for the answer. You are paying for the search.
Which is why “tokens got cheaper” and “the bill went up” are both true and neither one contradicts the other. Price per unit down by half. Units per job up five, ten, forty. That’s the arithmetic of the last eighteen months, and almost nobody has it on a slide.
The one genuinely aligned model in that census is per outcome — pay when the ticket actually closes. Salesforce moved its help agent to two dollars per resolution in June. Worth watching, because the fight there isn’t the price. It’s the definition. Who decides a resolution counts, and what happens when the customer reopens the thing eight days later.
Where I think most operators get this wrong: they read a rising bill as a vendor problem and go shopping for a cheaper model. So they downgrade, the agent fails more often, it loops harder, and the bill lands higher than where they started. Cheap tokens don’t rescue an undefined job. They just make it cheap to keep trying.
The bill is a diagnostic. If one job costs three times what you quoted, the machine is telling you the job was never defined tightly enough to be done. Vague brief. No stop condition. No test for finished. An agent authorised to keep going until it feels right. You didn’t buy intelligence. You bought an open-ended search with your name on the account.
So price the finished job and cap the thinking. Four things, none of them technical.
Put a ceiling on every agent: a step budget, a retry budget, a spend cap, and a rule that when it hits the ceiling it stops and asks a human instead of trying harder.
Count retries as defects, not effort. Every loop is a broken instruction, and fixing the instruction is cheaper than buying more tokens.
Route by task. Small model for the eighty percent that’s formatting and extraction, frontier model for the part that genuinely needs judgment. Gartner calls it inference tiering. The plain version is: don’t hire a specialist to sort the mail.
And keep your own number for what a completed job cost — per job, per client, per week. If you can’t produce that, you don’t have a unit economics problem. You have a visibility problem, and the only party who can currently see the answer is the party sending the invoice.
None of this is an argument against agents. It’s an argument for knowing what you’re buying.
Because the bad version of this doesn’t look like a bankruptcy. It looks like a business that can’t quote. The price to the client is fixed, the cost to deliver arrives thirty days later in a currency nobody in the room chose, and the only person who understands it works for the company selling it.
A price you can’t predict isn’t a price. It’s a wager, and the house is holding the meter.
So before the next pilot, two questions. Can you say what this job will cost before you start it? And can you explain the bill afterwards to somebody who wasn’t there when you built it?
If both answers are yes, the machine is leverage. If either is no, you haven’t bought a tool — you’ve signed up for a number you’ll be arguing about for a year.
You’ve got this. Just don’t let someone else hold the number.