Contents7
Two claims circulate, and both are true. AI has never been so cheap: at equal capability, the price of a model was divided by more than 280 in two years. And AI bills have never worried finance departments so much. The paradox is not one: it is the story of a unit price falling while volume explodes.
Our guide to AI costs in business explains how a bill is calculated and forecast. This article steps back to look at the market: what drives the bill up, and the unit that lets you judge it.
In short
The price of a given level of capability is collapsing: divided by more than 280 between November 2022 and October 2024, according to Stanford.
The volume consumed grows faster than the price falls: the total goes up.
Surprise bills hit one large company in two among those surveyed.
A large share of requests go to a model more powerful than needed.
The unit that matters is the cost per outcome (a request handled, a document produced), not the token price.
The unit price is collapsing
Stanford’s AI Index 2025, published by an academic institute with no commercial interest, tracks the public price of models at equal performance on a general knowledge test.
per million tokens in November 2022, at a given capability
for the same capability in October 2024
cheaper each year, depending on the task
A token is a piece of a word, the unit in which the text a model reads and writes is billed. At this pace, a task that was expensive two years ago costs cents today. If the story stopped there, AI budgets should be melting.
Volume grows faster
They are not. Bain, which sells strategy and AI consulting, has tracked both curves since December 2024, without a survey, from market data.
Token volume and cost per million tokens, index 100 in December 2024
The unit price lost more than half its value; volume more than quintupled. Multiply the two: total spending more than doubled (our calculation on Bain’s indices). Bain also quotes a bank, Wells Fargo, which found that higher usage has more than doubled the overall bill.
Definition
The rebound effect
When a resource becomes cheaper, people use more of it, to the point that total spending can rise. Economists also call it the Jevons paradox, after coal in the 19th century.
Where all this volume comes from
Three mechanisms push consumption up without anyone deciding to spend more.
- Uses multiply. Each price drop makes a use profitable that was not before: summarising every email, sorting every ticket, reviewing every contract.
- Agents chain calls. An AI agent does not answer in one call: it plans, calls a tool, reads the result, starts again. Each turn sends the whole context back, and is billed.
- Models are too powerful. According to Accenture, which sells consulting and managed AI services, 54% of AI requests go to a more powerful and more expensive model than needed, and 90% of the more than 9,000 tasks studied do not need a frontier model. The study covers 750 executives of companies above one billion dollars, far from our audience, but the mechanism does not depend on size.
Surprise bills
The result shows in the surveys, all conducted among large companies by firms that sell consulting and AI services.
of large companies surveyed by Capgemini received generative AI bills higher than expected
large companies surveyed by Accenture runs out of its token budget before year end
only of large US companies surveyed by KPMG see what their AI costs in real time
Capgemini’s study covers companies above one billion dollars, and Capgemini itself presents it as directional guidance; surprise bills hit companies in the pilot phase (57%) more than those scaling up (46%). KPMG surveys 204 companies of one billion dollars and more. At EY, which sells consulting and AI services, 82% of 534 US executives whose organisation already invests in AI say they are concerned about their token consumption.
The unit that matters: cost per outcome
A price per token does not tell you whether spending is good. A million tokens can produce a thousand useful answers or an agent going round in circles. The unit that lets you judge is the cost per outcome: what one processed supplier invoice, one resolved customer request, one translated and reviewed page costs.
BCG, which sells AI consulting, uses the same unit in its September 2026 report: the cost per successful outcome. Yet few companies know how to calculate it. According to Accenture, only 35% of the large companies surveyed can calculate a cost per outcome, and less than one dollar in five spent on tokens is tied to a quantified financial result.
Watch out
The opposite trap
Rewarding consumption encourages consumption. KPMG notes that 41% of the large US companies surveyed would consider rankings or bonuses based on token consumption, and sees the risk of rewarding activity rather than outcomes.
That is the unit we track on our own tools. On the translation pipeline we run for Palmora Property, a real estate portal in six languages, we count the cost per translated segment: switching providers divided it by 60, and 95% of segments come out of the cache after the first pass, without calling the AI again (project CNT.7). These ratios hold for that pipeline; they show what measuring per outcome makes visible.
Calculating your cost per outcome
For a team of 20 to 500 people, there is no need for a large-company dashboard to get started.
- Name the outcome.
A unit the business recognises: an invoice entered, a ticket closed, a product page published.
- Note one month’s spending.
On the provider’s console, per key or per project, for that use only.
- Divide.
The month’s spending divided by the month’s number of outcomes.
- Compare.
With the cost of the same outcome without AI, and with the previous month.
- Act on the gap.
A lighter model, a cache, batch processing, or stopping a use that brings in nothing.
The technical levers of step 5 are detailed in our guide to AI costs in business. And when AI comes bundled into software as credits, the same reasoning applies: what “included” AI really costs.
Key takeaway
Our position, signed Jim
The falling token price is good news that excuses nothing. An AI bill is judged by cost per outcome, use by use. We would change our minds if prices ever fell faster than volumes rise, durably.
If the bill is rising and you cannot say what it brings in, our AI at the right cost offer measures the cost of each use and gives each task to the model that is good enough.
Frequently asked questions
Why is my AI bill rising while prices fall?
Because the volume consumed grows faster than the unit price falls: more uses, agents multiplying calls, longer contexts. According to Bain’s indices, volume was multiplied by more than five while the unit price was divided by more than two.
What is the cost per outcome?
It is AI spending divided by the number of useful results it produced: invoices processed, requests resolved, pages translated. It lets you compare AI with the same work done another way, which the token price does not.
Should you always pick the cheapest model?
No: pick the cheapest model that does the task correctly. According to Accenture, most requests go to a more powerful model than needed; conversely, a model that is too weak and makes mistakes costs human rework.
How do you avoid a surprise bill?
By setting a cap per use with the provider, tracking spending every week and tying each access key to a specific use. Spending you cannot tie to any outcome is the first to cut.
Sources
- Stanford’s AI Index 2025 (new tab)hai-production.s3.amazonaws.com
- Bain (new tab)bain.com
- Accenture (new tab)accenture.com
- Capgemini’s study (new tab)capgemini.com
- KPMG (new tab)kpmg.com
- EY (new tab)ey.com
Jim · founder of Brain Plus AI
Also in business ai
Practical guide10 min
The cost of AI in business: tokens, routing and forecasting
What you really pay when you use an AI through an API: input and output tokens, the context sent back with every call, an agent's turns. And a method to forecast the bill before you receive it.
Choose and price11 min
AI credits in software: what "included" AI really costs
HubSpot, Salesforce, Microsoft, Odoo and n8n now meter the AI in their software in credits: an included allowance that runs out, then a rate card per action. We converted those credits into dollars to see what an answer, an enriched record or a resolved conversation really costs.
Stepping back8 min
Half of large companies profit from AI. What about the rest?
One study says nearly half of large companies finally get value from AI, another finds only 6% of leaders in the accounts, a third that 56% of CEOs see no significant benefit. Who should you believe, and what is a survey worth when everyone describes their own results?