AI Gets 13 Times Cheaper Every Year? For Whom, Exactly?

Don't feel like reading it all? The key points in a 4:33 video:

Watch on YouTube. The narrator's voice is AI-generated. Scenarios 2 and 3 are my own model calculations from Epoch's data. All sources are listed below in the article.

This week in my feed: "The same performance has gotten about 13x cheaper every year since 2023. No technology has ever fallen this fast, not electricity, not compute, not batteries. The direction is clear."

Below it, a chart. The AI line plunges straight down. Electricity crawls along for eighty years. The message is obvious: AI is the safe bet, and the data proves it.

The number comes from a study by Epoch AI. The study is good. The data is public. So I recomputed it.

Result: The number is right. The message isn't.

Let me be clear up front: this is nothing against Epoch. You can calculate it this way. You just have to put the number in context, and whoever reads it has to understand what it measures. Then "wow" turns into "wow, it depends."

The number in one sentence

Imagine someone says: "Flights get 13 times cheaper every year."

That sounds huge. Then you read the fine print. What they mean: the cheapest flight on any airline, at five in the morning, no luggage, booked by someone who searches every airline every quarter and rebooks on the spot.

That is exactly what the study measures. Not the price you pay. The cheapest price anyone on the entire market could pay, if they kept switching.

Epoch says so itself, openly: "Essentially no user stays permanently on the cost frontier." In plain words: practically no customer always buys at the cheapest edge.

That sentence is in the report. It is no longer in the LinkedIn post.

A world record can only get better

One more thing you need to know.

The 100-meter world record can only get faster. It can never get slower, because an old record stands until someone breaks it. That does not mean you run faster.

The 13-times-cheaper number is a world record like that. The study looks at the lowest price ever reached for a given level of performance. That price can only fall. So the fact that it falls is not a finding. It is built in. The only open question is how fast.

To be fair: I tested what happens if old models disappear after a year. Then the record could rise again. It barely does. New, cheap models arrive fast enough. At the cheapest edge, the price really falls.

The only question is: Are you standing at that edge?

Three customers, the same data

I ran three calculations on the same data. For three customers who actually exist.

The bargain hunter switches to the cheapest model every quarter, no matter who makes it. He always needs the same performance. This is the customer the study has in mind. His bill of €100 today costs €8 in a year, between 4 and 11 depending on the method.

The normal company buys from OpenAI, Anthropic or Google. Contracts, data protection, established workflows. It takes a typical current model and also always needs the same performance. In my calculation, its bill of €100 today costs €96 in a year, between 86 and 119 depending on the method.

Important: this is a model result, not a number from anyone's actual books. I take the model in the middle of all models the big three released over the past year that Epoch tested. A real company uses specific models for specific tasks, and how often each model actually gets used is in no dataset. The number shows the direction for someone who doesn't keep switching. It does not show what any particular company pays.

The demanding one takes every new model and turns it up to the highest setting. Her bill of €100 today costs about €330 in a year, 250 to 460 with statistical uncertainty. In return, her model can do a lot more than before.

The ranges matter. They show: the exact numbers wobble, the direction doesn't. The bargain hunter pays far less in every calculation. The normal company pays about the same in every variant of my calculation. The demanding one pays more in every calculation.

All three numbers come from the same data. None of them is wrong. The headline "13 times cheaper" only tells the first story.

Turn the dials yourself

I rebuilt the study's calculation. At the default settings it produces exactly what Epoch published. Then I turned every assumption you could reasonably make differently into a switch. Try it. If you want to follow every step, with formulas, data and scripts: method report as PDF (25 pages).

Eight mistakes on the way into the feed

1. "The same performance gets cheaper"

Yes. At the cheapest provider, if you switch immediately. That is the 5 a.m. budget flight.

Take a typical model instead, and in my calculation 13 times cheaper turns into 1.6 times cheaper per year. Stick to the big three providers, and almost nothing moves.

Most companies don't stand at the cheapest edge of the market. They stand right in the middle.

2. "13 times, every year"

That sounds like a train running at a steady speed. In reality it is more like an elevator that mostly stands still and sometimes drops five floors at once.

In most quarters, the price for a given performance does: nothing. Then a new model arrives, and the price crashes. The 13 is the average of many standstills and a few crashes.

Picture it: a millionaire walks into a bar. The average income in the room shoots up. Nobody at the bar earns a cent more. Take the normal quarter instead of the average, and almost nothing of the price drop is left.

And the crashes hinge on a handful of models. Take a single model out of the calculation for the chess puzzles, and the rate drops noticeably.

3. "Faster than any technology before"

Let's start with what Epoch itself writes about this. The report says, word for word:

"when comparing price drops for AI to drops for other technologies for which we have price series, we are comparing apples and oranges."

The authors say this themselves. In the feed it becomes: "No technology has ever fallen this fast." Why the comparison limps becomes clear once you draw it to the end.

The chart compares three years of AI with eighty years of electricity. That is like comparing a sprinter's start with a marathon runner's average pace. The sprinter always wins.

Many new technologies get cheaper especially fast in an early phase, then slower. So the only fair comparison is the fastest three years of each. Epoch even computed those numbers. They are in the Excel file that accompanies the study, not in the report. Why, we don't know. You can only speculate.

With a fair comparison, "4 times faster than DNA sequencing" shrinks to 1.3 times. "6 times faster than computers" becomes a little over 2 times. Fast, yes. Without precedent in history, no.

And then the part that surprised me most. Look at where the curves stop. The electricity curve ends in 1973. By then the price decline had already stalled for three years. In the fall of that same year, the oil crisis hit. The data after that is in the same Excel file Epoch published. I simply kept drawing.

For eighty years, electricity in the US got cheaper. For the fifty years after that, it didn't. By 1984 the price had risen 29%, then it slowly fell, and since 2002 it has been rising again. In 2025, electricity costs exactly as much in real terms as in 1973.

The UK looks the same. From 1900 to 1973, electricity there got twenty times cheaper. After the oil crisis it rose 30% by 1980, and in 2008 it cost as much in real terms as in 1973. In Western Europe, household electricity has become more expensive since 2007/08 after inflation: Germany up 26%, France up 61%, the EU average up 13%. This is not an American special case. This is the pattern.

And one more thing you only see if you look closely: on Epoch's scale, you can't see a rise at all. The axis spans twelve orders of magnitude, from the starting price down to a trillion times cheaper. Above the starting price, there is no room at all. A 30% price rise is a few pixels on it. Switch the view in the chart above. On "Ordinary scale" you see what happened after 1973. And in "Pace per decade" you see that every one of these technologies had fast phases, slow phases and setbacks. Same data, three pictures. Which one goes in the report is a choice. Not a measurement.

This is not an accusation against Epoch. The chart counts years since the start of the price decline. It deliberately shows only the fall. But that is exactly the lesson: every one of these technologies had a phase in which it got cheaper at breakneck speed. And for electricity, that phase ended when the world outside changed. Oil got expensive, and suddenly the direction was no longer clear.

On top of that: for electricity, what counts is the price customers paid on average. For AI, it is the cheapest price on the entire market. Those are exactly the apples and oranges Epoch is talking about.

Take the train test. Deutsche Bahn, Germany's national railway, advertises ICE high-speed tickets from €8.99 (source, German). The offer is real. But it is limited, and hardly anyone pays that price. Certainly no business traveler: you travel when the meeting is, not when the saver fare applies. Nobody would conclude from it that train travel has become rapidly cheaper. One single, limited offer is cheaper. That's all it says. It doesn't apply to any other ticket. Statistically, a price like that is an outlier: a small allotment at the edge, far from what a trip costs on average or typically. If you want to know what train travel costs, you sort it out as noise. The study does the opposite. It measures only the edge.

And while the saver fare stays put, the regular price climbs. With the December 2024 timetable change, Deutsche Bahn raised its flexible full fare by 5.9% on average, while the Super Saver fares stayed stable (source, German). As the reason, it cited significantly higher costs, especially from the latest wage agreements (source, German), and two years earlier, energy costs (source, German). To be fair: in December 2025, Deutsche Bahn skipped a long-distance price increase (source, German). The pattern holds: the teaser offer stays cheap, the price most people pay goes up. And the cost drivers are staff, power, materials and maintenance. A data center depends on those too, power and chips above all.

Same with flights. A domestic flight in the US has gotten 37% cheaper in thirty years after inflation, on average from $618 to $387 (source). A little over one and a half percent per year. If you instead tracked, like Epoch, the cheapest promotional fare any airline ever sold, you would get a spectacular number. Only nobody who has to fly pays it. Back in the early 2000s, according to American Express, a typical business fare already cost five to six times a leisure fare.

That is exactly how the sentence "No technology has ever fallen this fast" is built. It compares AI's promotional price with the average price of everything else. By that method, the flight would also be the fastest technology of all time. Epoch says this itself, by the way: "we are comparing apples and oranges".

4. "Performance" means work

The measuring sticks are a math olympiad, research mathematics, exam questions in physics and chemistry, chess and a secret game. Good tests of reasoning. But no company makes money solving chess puzzles.

The two tests in the dataset closest to everyday work: coding and factual knowledge. That is where the price falls the slowest, a little over 5 times instead of 13 times per year. For the typical model of the big three providers, it even rises slightly there.

And at which level? Epoch averages across all performance levels, from "barely a task right" to the top score. You might suspect the 13 comes from the trivial levels. I checked, and it doesn't. For high performance, 75% correct and up, the price actually fell faster: about 30 times per year. That speaks for Epoch. But only at the cheapest edge. For the normal company that sticks with the big three, the same high performance gets a mere 1.2 times cheaper per year. The better the performance, the bigger the gap between headline and bill.

With three catches. High levels were only reached recently and make up less than a tenth of the data points. So what gets measured is exactly the moment a new top score loses its premium. On the math olympiad, the ceiling has been hit, 100%, it can't get better, only cheaper. And on the secret game, the one test nobody can train for, the best models still solve at most 30% of the tasks.

5. "You pay a little more and get many times the intelligence"

The study never measured that. It holds performance fixed and looks at what it costs.

And here sits what may be the biggest problem, bigger than everything before. It affects every number in this article, my own three customers included. They all hold performance fixed. For a price comparison, that is correct. Methodologically, Epoch does nothing wrong. It only becomes a problem when it turns into a forecast of what AI will cost a company. Because nobody buys the same performance forever.

In 2023 it was impressive when a model wrote a usable text. Then it was supposed to analyze documents and write code. Then operate tools on its own and solve multi-step tasks. Today many people expect entire workflows. The benchmark performance of 2023 now costs almost nothing. But who today wants to buy only the performance of 2023?

A stagecoach managed about 10 kilometers per hour. The ICE, Germany's high-speed train, runs at up to 300. A coach ride would be dirt cheap today. Only nobody takes one anymore. The coach didn't get more expensive. Faster travel became possible, and because it became possible, everyone used it. The requirements for travel went up. What becomes possible gets used. And paid for.

We measure the price drop against yesterday's requirements and sell it as tomorrow's savings.

Even if the 13 were absolutely correct under every assumption, it would not follow that companies will pay 13 times less for AI. The decisive question is not only how cheap the same performance gets. It is which performance will count as good enough tomorrow.

If you turn every new model up to the highest setting, you experience the opposite of cheaper. The new models think longer before they answer. That costs money. On every single test in the data, thinking effort grows from year to year. The bill per task rises several times over per year on most tests.

That is why the people who say "My AI bill is going up" are right. And the study is right too. They just mean different things. Whether the extra is worth it is not decided by the price. It is decided by whether the additional performance does anything in your business.

6. "The direction is clear, so don't wait"

This is where a measurement turns into an investment tip. And the data doesn't support it.

New top performance is expensive at first and then gets cheap fast. According to Epoch, its price initially falls by about two thirds per quarter. If you only need a capability cheaply, a year of waiting gets it for a fraction. The same number that is supposed to prove "get in now" argues just as well for waiting.

And for the providers themselves, Epoch says profits on a model are "fleeting". A market in which performance loses value this fast is no safe bet for investors. It is a risk.

Starting early can still be right. Because you learn which process fits, where the data lives, who checks the result. That is a good argument. It just doesn't need the 13. Propped up by a number that measures something else, it gets weaker, not stronger.

7. "It'll keep going like this"

Epoch measures yesterday's list prices, meaning what providers charged. The study does not separately account for power, chips and investor money. They only show up indirectly in those prices. Whoever turns that into "the direction is clear" projects it into the future. And silently assumes: the world stays as it is.

Three things inside an AI price are moving right now. Power. Chips. And investors' money.

Power. In the largest power market in the US, the wholesale price rose 76% from early 2025 to early 2026. The independent market monitor attributes this mainly to data centers (source). So AI is driving up the price of its own raw material.

Chips. Renting an AI chip on an annual contract got 38% more expensive from fall 2025 to spring 2026 (source). Memory chips cost almost twice as much in early 2026 as a quarter earlier (source).

Investor money. Part of today's prices doesn't cover costs. According to the Wall Street Journal, OpenAI had $6.7 billion in revenue in the second quarter of 2026 and sank even deeper into the red. Anthropic, by contrast, reached an adjusted operating profit in the third quarter of 2026, according to the Financial Times. So the situation differs by provider. Only one thing is clear: whoever sells below cost does so as long as someone pays the difference.

And these are not last year's numbers. The calculator above shows the actual trends of the last five years. Wholesale power on the German exchange cost a good 70% more in September 2026 than a year earlier. Renting an AI chip went up 28% in twelve months, almost 9% in the last week alone.

And now the part that comes from the data itself: I broke the price drop apart. About three quarters of it came from cheaper tokens, only a quarter from models getting by with fewer tokens. So the big part hinges on power, chips and pricing policy. If token prices stop falling, the bargain hunter no longer gets 13 times cheaper, but 1.8 times. €100 then turns into €57 instead of €8.

In the calculator above, you can change the world. Press "All of it at once". Then even the bargain hunter pays more in a year than today. Is that likely? Maybe not. Impossible? No. And that is exactly the point: "The direction is clear" claims it is impossible.

8. "If it's going this well, it's a sure thing"

A simple question: if AI were so mature and so safe, why do the providers rush the next model onto the market every few months?

The answer is in Epoch's data. New top performance loses its price faster than almost any other product, at first by about two thirds per quarter. If you don't keep shipping, your model is bargain-bin stock within a few quarters. Epoch itself calls the profits per model "fleeting".

That is not a sign of strength. That is a treadmill. Stand still and you fall off the back. And if you shop on that treadmill as a customer, you buy something that is already old tomorrow.

Not just me

I'm not alone in seeing it this way.

Toby Ord, philosopher and AI researcher at Oxford, showed back in 2025 that the progress of the new models comes mainly from them computing more when they answer. His conclusion, about the economics of the AI companies: "So if it costs a factor of 10 or 100 times as much to answer each question, this really does affect their bottom line. And unlike training costs, these costs can't be made up in volume." (source) That is exactly what our data shows for the third customer, the one who turns every new model all the way up: the models think longer and longer per task, and her bill rises even though every single token gets cheaper. Epoch takes Ord seriously: the report explicitly thanks him in a footnote for his objection to the technology comparison. And Epoch answered him publicly: the price for a fixed performance still falls fast (source). To be fair: Redwood Research considers Ord's method too pessimistic (source).

Gartner predicts, in a forecast from August 2026, that the cost per agentic AI workflow will rise more than fivefold through 2028, even though tokens get cheaper. Gartner calls this the "Inference Paradox" (source).

A second research group led by Gundlach measures much like Epoch and lands at 5 to 10 times cheaper per year, not 13. Epoch cites the work itself. Even the researchers don't agree on the size.

And Pasquale Pillitteri writes on his blog, in essence, what drove me to write this article: the study's caveats get lost when it is shared, and in a company's books there is no such thing as fixed performance. If you plan, plan with cost per task, not price per token (source).

None of us knows how AI prices will develop over the next few years. I don't either. But the data does not support a clear conclusion of "it's getting dramatically cheaper." What it supports: it depends, dramatically.

What holds

So this doesn't come across wrong: the price for a fixed AI performance is falling. Fast. I tried 512 different, defensible ways to calculate it. The vast majority point down.

But the range is huge. The same data yields anything from 26 times cheaper to 1.3 times more expensive per year. 53 of the 512 calculations show rising costs. That is not a probability, because the calculations are not all equally plausible. It only shows how far you can get with defensible choices.

Between 26 times cheaper and more expensive, there is no rounding error. Between them lies the question of who you are, what you use AI for and how you buy.

And all those formulas?

The report contains a lot of formulas: logistic curves, Box-Cox transformations, envelope curves, stochastic frontier models. It looks like proof. Do you need any of it for the 13? No.

The 47% per quarter comes from a simple calculation with no formula at all: cheapest price today versus cheapest price a quarter later, averaged across all levels. That is exactly what I rebuilt. Epoch's preferred formula model lands on the same number.

So what are the formulas for? They smooth the staircase into a surface. That lets you make statements like "right after a new top score, the price falls by two thirds per quarter." And that is exactly where it gets shaky: on 40 to 63% of the surface the model colors in for this, there are no measurements at all. There the formula just extrapolates.

And Epoch's own table shows something else. The same data with a different formula model, fitted to all models instead of only the cheapest edge, gives 5 to 10% per quarter instead of 47. That is 1.2 to 1.5 times cheaper per year, pretty much exactly our normal company. Epoch rejects this variant, for understandable reasons: it is not built on the cheapest edge. But it shows what this whole article is about. A formula is not a measurement. It is one more assumption.

Six questions for any price statistic

They work for AI, for electricity and for anything someone shows you with an arrow pointing down.

1. Whose price? The cheapest on the whole market, a typical one, or yours?
2. For which performance? Measured against what, and does it have anything to do with your work?
3. Average or normal case? Does the number move evenly, or is there a millionaire in the bar?
4. Could the number even come out differently? A world record that can only fall doesn't prove that anything is falling.
5. Compared with what? Sprinter against sprinter, or sprinter against marathon runner?
6. What has to stay the same for this to continue? Power, chips, investor money. And who guarantees that?

If you can answer these six questions, you don't need the headline anymore. If you can't, you shouldn't believe it. Mine included.

For nerds: how I calculated

Every step, with formulas, data and scripts, is in the method report (PDF, 25 pages). Here is the short version.

The basis is Epoch AI's data and code, as of October 2026 (github.com/droodman/inference-cost). I rebuilt the study's model-free calculation in Python. It reproduces Epoch's table exactly: 47.1 / 43.0 / 53.1 / 47.0 / 44.0% per quarter for the five core benchmarks. I did not rebuild the smoothed regression models. On average they land on the same number.

The switches in the calculator and their effect, on Epoch's five benchmarks:

More switches under "For statisticians": the grid across performance levels (Epoch's own Figure 7 shows four variants, 9.5× to 23× per year), the measurement period, simulated versus real runs, model lifetime and corrected release dates.

"Typical model" means: the median across all models released in the last twelve months, each at its cheapest setting that reaches the level. That depends on which models Epoch tested. It is a second reading, not a better truth.

"The demanding one" uses each model's strongest setting, with no truncation, and the trend of her cost per task across release dates, as a geometric mean over the five benchmarks: about 3.3× per year. Tokens per task grow by 1.5× to 5.6× per year, depending on the benchmark.

The split into "fewer tokens" and "cheaper tokens": at every data point, cost equals tokens times price per token. I compare the record holder now and one quarter later and split the change between the two factors, averaged as Epoch does. Result across the five benchmarks: three quarters cheaper tokens, one quarter fewer tokens. On coding and research mathematics, the record holders actually need more tokens than their predecessors. "Cheaper tokens" also includes smaller, more efficient models, not just outside factors. Freeze the token price, and on average 1.76× cheaper per year remains.

The world scenarios in the calculator are not a measurement but assumptions. The conversion uses rough cost shares of an AI data center: chips and servers a little over half, power about a seventh. There is no clean public number for this. The default values (power up 50%, chips up 40%, price mark-up 50%) are based on the 2025 and 2026 developments linked above.

How robust are the three customers? I checked. The normal company: whether the "typical model" is determined over 12, 18 or 24 months changes little (€93 to €96). Leave out each of the 75 models one at a time, and the bill lands between €86 and €109. Only with a six-month window is it €119. The demanding one: a 90% bootstrap over the models gives 2.5× to 4.6× more expensive per year. Important: that applies to "every new model at the highest setting." For the record holders alone, there are too few models to say anything (bootstrap 0.5× to 2.9×). The bargain hunter: the range across Epoch's own four calculation grids from their Figure 7.

What the 512 calculations are not: a sample. They are neither independent nor equally likely. I count them only to show the range, not to claim a probability. Real confidence intervals for Epoch's core number don't exist, not at Epoch either. The study explains why at length: a record price is an extreme value, and the usual bootstrap doesn't work for it.

Robust: dating, correction for lucky guesses, simulated budgets (real runs actually show a faster decline) and the assumption that old models stay available. Shaky: which price, which providers, which tasks, average or median.

A side note for Epoch: the code assigns release dates by matching the beginning of the model name. As a result, o3 gets the date of o3-mini (January 31 instead of April 16, 2025), o1 gets o1-mini's, QwQ-32B gets the preview's. The correct dates are in Epoch's own table. The effect on the headline number is minimal.

Related

• AI: more expensive or cheaper? 26 positions, no consensus (German)
• 76 Percent Is Not an Answer
• Explaining Is Easy, Predicting Is Hard
• Fear Is Not a Probability
• The Disappointment Was Priced In