Cheaper Never Meant Less

In November 2021, a million tokens of GPT-3-class intelligence cost $60. Three years later, a model matching it on the standard benchmark charged six cents — a thousandfold collapse, roughly ÷10 every year, and that is the slow lane: track the price of hitting a fixed capability across many tasks and the declines run anywhere from 9× to 900× a year, with a median around 50×. Show that curve to any CFO in 2023 and ask what the company's AI budget looks like after three years of it, and the answer writes itself: a rounding error.

The budgets did the opposite. Google's models chewed through 9.7 trillion tokens a month in the spring of 2024. A year later it was 480 trillion. This spring: 3.2 quadrillion — a meter running roughly 330 times faster than two years before, over the very stretch in which the thing being metered got orders of magnitude cheaper. The pattern — unit price collapses, total spending rises — feels like it should be a scandal, or at least a mistake. It is neither. It is one of the oldest, best-documented patterns in industrial economics, and it was named a hundred and sixty years ago, about coal.

A Victorian economist already ran this experiment

In 1865, Britain was arguing about whether its coal would run out, and the comforting answer in circulation was: don't worry — engines keep getting more efficient, so we'll burn less. The economist William Stanley Jevons wrote The Coal Question largely to demolish that comfort. "It is wholly a confusion of ideas," he wrote, "to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth."

His evidence was the steam engine itself. James Watt's design used a fraction of the coal of the Newcomen engine it replaced — and Britain's coal consumption did not fall. It exploded. Because efficiency is not a reward for existing use; it is a price cut on the useful unit — the pumped gallon, the mile hauled, the lumen, the answered question. And a price cut does two things at once. It saves money on everything you already do — and it makes profitable a thousand uses that were not worth the fuel before. The efficient engine left the mineshaft and went into mills, boats, and locomotives. The second effect buried the first.

Seven centuries of cheaper light

The cleanest long-run record we have of this is light. Economists reconstructed what a unit of artificial light cost in Britain across seven centuries — candles to gas to electricity. By 2000, lighting services cost less than one three-thousandth of what they had in 1800. Did people keep their 1800 allotment of light and pocket the difference? Per-capita consumption rose 6,500-fold; total consumption, 25,000-fold. Every generation took the price cut and spent it on more light — brighter rooms, lit streets, lit nights — until an amount of illumination that would have bankrupted a Georgian household became too cheap to think about.

Efficiency sets the price of the unit. Appetite sets the size of the bill.

The arithmetic your gut refuses to run

Strip the paradox and there is no paradox — there is a multiplication. A bill is price × quantity. Efficiency moves the first factor down; the open question is what the second one does, and your intuition quietly assumes the answer is nothing. That assumption has a name — economists call the response of quantity to price elasticity — and a threshold: if a tenfold price cut grows your usage by less than tenfold, the bill falls, exactly as the gut expects. If usage grows by more than tenfold, the bill rises, and the price cut is the thing raising it. That single tick — ×10 more use per ÷10 of price — is the entire border between the world where efficiency saves you money and the world where it costs you. The mistake isn't bad math. It's doing subtraction where the world is doing multiplication.

Tokens are the new coal

The week a Chinese lab's cheap model wiped hundreds of billions off American tech stocks, Microsoft's CEO reached for the 1865 reference by name. "Jevons paradox strikes again!" Satya Nadella wrote. "As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of." Self-serving, yes — he was talking his own book. But the meter readings above say the book was right so far: the market's appetite has been growing faster than the price has been falling — usage up ~50× in a year when constant-capability prices fell ~10× — which is precisely the regime where cheaper means bigger bills.

Two engines drive the appetite, and both are worth naming, because both will show up on your own invoice. First: demand doesn't shop in the bargain bin. The tenfold discount is real — but it applies to last year's capability. The frontier model, the one you actually want, costs roughly what the frontier model always cost, and the moment a better one ships, demand migrates to it. Second, and larger: the size of an ask exploded. A 2023 chat reply was a thousand-ish tokens. Then models learned to think in tokens before answering, and to work in loops: a single "deep research" run burns about a dollar of inference; an agent left working for a day, on the order of seventy. The flat-rate "unlimited" subscriptions built on 2023-sized asks met 2025-sized asks, and the most famous $200-a-month one was rolled back within months. Cheaper tokens didn't shrink anyone's bill. They made enormous asks feel affordable — and the asks obliged.

The honest doubt

Now the part a site like this owes you: Jevons is a pattern, not a law, and the strong version of it is rarer than the current discourse suggests. When energy economists actually measure the rebound — how much of an efficiency gain gets eaten by extra use — the typical answer is 5 to 30 percent, not 100-plus. Full "backfire," where efficiency raises total consumption, has almost no well-documented modern cases in energy. Make your home's lightbulbs efficient and your lighting bill really does fall, because your appetite for light is, at last, nearly full. The paradox lives only where appetite is open-ended — where the price cut keeps unlocking uses that were unthinkable at the old price. Light in 1800 was such a frontier. Compute has been one for seventy years. Tokens are one now — which is exactly why the burden of proof sits on anyone forecasting that this price collapse, uniquely, ends in smaller bills.

And hold even the pattern loosely. Jevons saw the multiplication perfectly and still read the wrong future off it: The Coal Question forecast Britain's decline as its coal ran down, and instead the economy swapped fuels and moved on. Appetites saturate; substitutes arrive; a trend of ÷10 a year is a fact about the past three years, not a property of the universe. The claim here is narrow: when the unit price of a general-purpose thing collapses, "we'll spend less" is the one forecast with centuries of evidence stacked against it.

How to read a falling price

We built the multiplication so you can run it yourself: The Jevons Machine. Two dials — how fast the price falls, how your appetite answers — and the whole paradox turns out to be one tick on one of them: below ×10, efficiency saves you money; above it, efficiency is what's raising your invoice. Its natural companion is Your Number Was Never a Number — because "tokens get cheaper, so we'll spend less" is exactly the kind of confident point estimate that was always hiding a cloud.