Model maxing: why your AI bill grows faster than its value

What model maxing means

Which AI model is your business using today, and who chose it?

A model is the engine behind an AI tool: the software that reads a request and writes a response. Providers sell several, at different sizes. The largest are the most capable and the most expensive, often by a wide margin. Model maxing is my name for the habit of choosing the largest one for every task, just in case. A meeting summary, a tidied spreadsheet, a routine email: all go to the engine built for the hardest problems.

You pay for AI by the token. A token is a small chunk of text, a short word or part of a longer one, that the model reads or writes. Larger models charge more for each one. Choose big by habit and the bill follows the habit, not the need.

Why sensible people default to the biggest

I think this is predictable human behaviour, and the research behind it is long established.

Fear of missing out is the worry that others are having rewarding experiences without you. Przybylski and colleagues set out how to measure it in 2013. It was studied in social media, but I see the same pull in business: if a rival says it runs on the newest model, nobody wants to be on last year's.

Loss aversion comes from Kahneman and Tversky's work on prospect theory in 1979: people feel a loss more sharply than an equal gain. Choose a cheaper model and it fails, and the failure is yours. Choose the best and it fails, and nobody asks why. There is an old saying in technology buying that nobody was fired for choosing the biggest supplier. The AI version writes itself.

Defaults matter too. Samuelson and Zeckhauser showed in 1988 that people stay with the pre-selected or current option more than economic theory predicts. When a tool opens with the largest model already chosen, you would expect many people to leave it there.

Then there is herding. Banerjee, and separately Bikhchandani, Hirshleifer and Welch, each modelled in 1992 how people copy the choices of those before them, even against their own information. Provider league tables ranking models give that instinct a daily supply of evidence.

Underneath all of this sits a plain gap. The person choosing the model rarely sees what it costs. The bill goes to finance, weeks later, as one line.

What it does to the bill

Ramp, a spend management company, publishes an AI Index drawn from its own customers' card and billing data. It is vendor data and reflects Ramp's customer base, so I treat it as a signal and not a UK benchmark. Between January 2025 and April 2026, Ramp reports that token use among businesses with connected AI grew by 1,001%. Total spend grew by 497%.

Spend rose more slowly than usage, which means each token got cheaper. The total still climbed. And the mix moved upwards. In April 2026 Ramp found premium models accounted for 45.8% of tokens but 55.9% of cost. In June 2025 their share of cost was 5.7%.

Some of that shift is sensible. Certain work does need the strongest model. My reading is that a move from 5.7% to 55.9% of cost in under a year is also what drifting defaults would look like. Without task-level data, no one can say which it is. That is the problem.

The costs that never reach the invoice

The visible bill is the smaller part of the story. These are the ones I would look for first.

• Context bloat. The context window is the amount of text a model can hold in view at once. Pasting in whole documents and long histories to be safe means every request carries the extra weight, and you pay per token for all of it.
• Repeated resends. In many setups a long conversation sends the earlier exchange again with each new question. The same text is bought many times.
• No owner of the total. Spend sits across team cards, personal subscriptions and one-off keys. Everyone owns a slice. Nobody owns the sum.
• Forecast misses. A budget set from last quarter's usage is out of date by the time it is approved, because the tools and defaults keep changing.
• Shadow use. When the approved route is slow, people use personal accounts. That leaves no record, no rules and no way to count the spend.

A better default

The best models have a place, and the aim is to choose them on purpose. I would sort the response into three moves.

First, see it. Know which tools and models are in use, by whom, and what each costs. Second, match the model to the task. A smaller model can sort, label and draft routine work. A larger one earns its price on judgement and complexity. A person should decide anything with real consequences. Sending every parcel by same-day courier is a fair picture of the alternative. Third, put rules and an owner around it: who may use which tier, what the budget is, and who reviews the exceptions.

I will not quote a saving here. It depends on your tasks, and it should be tested against your own work. At GenWA, the place we start is Analyse, delivered through our advisory service. The health check is a short questionnaire that shows where you stand, and the cost saving calculator lets you test your own assumptions. Neither requires you to change anything first.

What to do on Monday

Five things you can do without buying anything.

• Ask who chose your current AI model for each main tool, and on what basis.
• List every AI subscription, card and key in the business, including personal ones used for work.
• Name one person who owns the total AI spend, not a slice of it.
• Pick one routine task and have it tested on a smaller model, with a person judging the result.
• Ask your finance lead for last month's AI cost by team, and note how long the answer takes.