Trend 1: the 2023 to 2024 pace
Spending in early 2026 was 0.49 times this trend.
Test: carried forward, this trend predicts about double what Anthropic actually spent in early 2025, before the bend.
1. The puzzle
In November 2026 two sets of numbers about one AI company met on X, and a quick conclusion followed. This page works through that conclusion slowly. It looks at the numbers four ways, and each one teaches something about how economists think.
Here are the numbers.
Spending rose 5.3 times, progress 1.9 times: each dollar bought 63% less progress
The money. Anthropic makes the Claude AI models. In the first half of 2025 it reported $1.85 billion of research and development expense. In the first half of 2026 it reported $9.74 billion. That is 5.3 times as much.
The progress. METR is an independent group that tests AI models. It scores each model on one scale and draws a line through the scores. Until October 2025 the line rose 15 points a year. After that it rose 29 points a year. That is 1.9 times as fast.
The sum. Divide progress by money. On METR's fitted lines, each billion dollars in early 2025 came with 4.05 points of progress. In early 2026, each billion came with 1.49. Progress per dollar fell by 63%.
So: five times the money, twice the speed, and each dollar buying far less. It looks like a field running out of easy wins.
One thing you should know before you go on. Much of the work on this page was done by Claude, and Anthropic makes Claude. Read it with that in mind, and check our sums. The code is open.
Margin notes carry a tag. MEASURED: read from a filing or dataset. QUOTED: what a named person said. CLAIMED: a finding by others that we have not reproduced. ESTABLISHED: standard arithmetic or method. OUR GUESS: our judgement. ILLUSTRATION: a drawing, not data.
Séb Krier writes about AI and policy. On 7 November someone asked him what would convince him that AI is now improving itself. In his answer he pointed at the two numbers above: "a >5x jump in spend for a ~1.9x faster benchmark slope". His reading: "we're already seeing diminishing returns".
Michael Wiebe, an economist, then did the division and drew the chart: "progress per R&D dollar fell by 63%, even while progress per year doubled."
Two days later Tyler Cowen linked the exchange on his blog under the title "Seb Krier, Ricardian?" David Ricardo was one of the economists who noticed, two hundred years ago, that if you keep adding workers to the same field, each new worker adds less grain.
Their arithmetic is right. We checked it against the filing. Wiebe's chart also carried a warning: "a fall in points per dollar does not establish that AI failed to improve research productivity."
That warning is the subject of the rest of this page.
A number means little alone. So ask what progress per dollar would have done if the speedup had never happened.
Showing pace doubled.
Spending rose 5.3 times and progress rose 1.9 times. Each dollar bought 63% less progress.
Why?Look at what the switch shows. With no speedup, progress per dollar falls 81%. With the speedup, it falls 63%. So most of the fall comes from the extra spending, and the speedup made it smaller. The question that matters is whether that spending would have happened anyway, and what it bought.
One fact about the per-dollar sum changes how to read the chart. It does not use how much progress happened in each quarter. METR took the scores from before October 2025 and drew the best straight line through them, then did the same for the scores after. The per-dollar sum uses only how steep each line was: 15 points a year, then 29. (METR's chart also shows each model's score; section 5 returns to those.) So within each period, the progress figure is the same every month. It was set that way when the line was drawn.
Think of a mango tree. I tell you it grew 30 centimetres a year for five years, then 60 centimetres a year after that. That is all I tell you. Now you ask: how much did it grow per bucket of water each month?
You can do the division. But my straight lines have already fixed the growth at the same amount every month. So your answer goes up and down only with the number of buckets. In a month with twice the water, "growth per bucket" halves. That tells you about the watering. It tells you nothing new about the tree.
Progress per dollar works the same way here. Inside each straight line, it moves only when spending moves. Double the spending and it halves. Spending rose fast, so it fell fast.
There is a deeper problem, and it would remain even if I had measured the tree every month. Water is not the only thing that changes. Sunshine, rain, soil, wind and the season change too, and you cannot hold them still while you count buckets. You may not even know the full list. So "growth per bucket" mixes the water's effect with everything else that happened that month. AI research is the same: money, staff, cheaper computing, better tools and new ideas all moved at once. Sections 4 and 5 come back to this.
For now, there is one narrower thing the lines can show: the moment they bend. The bend, October 2025, is the date where METR's two lines fit the scores best. It was chosen afterwards, from the data. It is not the day AI tools arrived: Anthropic used AI tools in its research before then. So the question becomes: what was spending doing at the bend?
A family's grocery bill has doubled every year for three years. This year it doubled again. Did something change in that house this year?
Not necessarily. The bill did what it had done for three years, but that does not prove nothing changed: prices could have risen while the family bought less. What the old pattern gives you is a guess at what the bill would have been. If the pattern said "double" and the bill tripled, the gap is worth explaining.
Economists call the old pattern carried forward the trend. Comparing with the trend is a common way to guess what would have happened anyway. It works only if the old pattern was steady and would have carried on.
Anthropic's research spending was not sitting still before the speedup. The filings report it over stretches of different length: whole years, half-years and nine months. To compare like with like, the table shows each stretch as an average per quarter, that is, per three months.
| Period | Average spending per quarter | What the filing reports |
|---|---|---|
| 2023 | $0.11 billion | $0.42 billion for the year |
| 2024 | $0.57 billion | $2.28 billion for the year |
| January to June 2025 | $0.93 billion | $1.85 billion for the half-year |
| July to September 2025 | $1.74 billion | worked out from two reported totals |
| October to December 2025 | $2.74 billion | worked out from two reported totals |
| January to June 2026 | $4.87 billion | $9.74 billion for the half-year |
Anthropic's later filing also gives July to September 2026: $9.50 billion in that one quarter. This page keeps the first half of 2026 as its end point, because that is the period the original exchange compared.
Now look at how fast the bill was growing, and notice that the speed itself kept changing. From 2023 to 2024, spending rose about five times. From 2024 to the first half of 2025, it was roughly doubling each year. Then in mid-2025 it took off: from the first half of the year to July to September, spending per quarter nearly doubled in four and a half months.
So the old pattern was itself shifting just before the bend. That makes "the trend" hard to pin down.
There is no single right way to carry the old pattern forward. None of the people in this story used a trend: Krier and Wiebe compared with last year, and Garicano and Imas, in section 5, fitted a model. The six below are ours, the obvious ways a careful person might carry the past forward. Each can be defended.
In each chart, every step up the side is ten times the one below. On a scale like this, spending that grows at a steady rate draws a straight line. Each trend is fitted to spending totals over their full periods.
Short line: average spending per quarter over the period it covers. Dashed line: the fitted trend. Red line: the bend, October 2025. Dotted connector: the latest period's average compared with the trend over that same period.
The number under each chart compares spending in early 2026 with the trend averaged over that same period.
Trend 1: the 2023 to 2024 pace
Spending in early 2026 was 0.49 times this trend.
Test: carried forward, this trend predicts about double what Anthropic actually spent in early 2025, before the bend.
Trend 2: the 2024 to early 2025 pace
Spending in early 2026 was 2.70 times this trend.
Test: this trend predicts about a third less than Anthropic spent in July to September 2025, before the bend.
Trend 3: the 2025 pace before the bend
Spending in early 2026 was 0.91 times this trend.
Test: this trend matches the totals for its two periods, because it was built from them. That proves nothing.
Trend 4: 2024 to 2025 yearly growth, anchored to early 2025
Spending in early 2026 was 1.89 times this trend.
Test: this trend uses figures from after the bend, so the test does not apply. It predicts about $4.94 billion for all of 2025, against $6.34 billion actual.
Trend 5: the same yearly rate, starting from the last quarter before the bend
Spending in early 2026 was 1.46 times this trend.
Test: this trend uses figures from after the bend, so the test does not apply.
Trend 6: first half of 2025 to second half of 2025
Spending in early 2026 was 0.89 times this trend.
Test: this trend uses figures from after the bend, so the test does not apply.
Trends 1, 2 and 3 use only figures from before the bend, and they disagree wildly. Measured against them, spending in early 2026 was anywhere from 0.49 to 2.70 times what the old pattern predicted. Trends 4, 5 and 6 use some later figures, which muddies them.
So the honest answer to "was spending unusual at the bend?" is: it depends on which past you treat as normal. Take the year and a half before the bend as normal, as trend 2 does, and spending raced ahead of its trend about one quarter before progress sped up. That fits Krier's reading: more money first, then more progress. Take only the few months just before the bend, as trend 3 does, and spending looks ordinary.
Take trends 1 and 2. Both are built only from figures before the bend, yet one says spending fell far behind and the other says it raced ahead. Here is where their two numbers come from.
Each number is actual spending in early 2026 divided by what the trend predicted. Above 1, spending ran ahead of the old pattern. Below 1, it fell behind.
Trend 1 copies the pace from 2023 to 2024, when spending rose about 5.4 times in a year. Carried forward to early 2026, that pace predicts about $10 billion a quarter. Anthropic actually spent $4.87 billion a quarter. Actual divided by predicted is 0.49. By this yardstick, spending fell well short.
Trend 2 copies the pace from 2024 to early 2025, when spending rose about 1.95 times a year. Carried forward, it predicts about $1.80 billion a quarter. Actual divided by predicted is 2.70. By this yardstick, spending raced ahead.
Same spending, two yardsticks, opposite verdicts. There are two reasons they differ so much.
First, Anthropic's spending growth kept changing. It rose about five times from 2023 to 2024, then roughly doubling each year into early 2025, then surged in mid-2025. Each trend copies a different stretch of that history.
Second, growth compounds. Trend 1's rise of about five times a year, carried from the 2024 average to the early 2026 average, multiplies spending more than fifteen times. Trend 2's rise of about 1.95 times a year, for one year, multiplies it by about 1.95. A gap between two yearly rates, run over different lengths of time, becomes a huge gap between two forecasts.
Think of measuring a child's height. Measure during a growth spurt, carry that rate forward, and you forecast a giant. Measure during a quiet year and you forecast someone short. Both forecasts use real measurements. Neither tells you the adult's height.
So a better question than "which trend do I like?" is "did this trend predict the figures just before the bend?" The test line under each trend in the panel answers that. No trend passes cleanly.
The honest conclusion is that the past gives no steady trend to compare with. That, more than any single number, is why the money-versus-AI question is so hard to answer.
"Five times the money" is a fair description of what happened. It cannot, by itself, say what caused the speedup. For that you would need to know how much more money went in than would have gone in anyway, and on these figures nobody can say, not even whether it was more or less than usual.
Suppose you could settle on a trend. You still could not say how much of the speedup came from money and how much from AI. Astra, an AI model from a different company, reviewed our reasoning and showed why with a small example.
A kitchen's budget doubles, and it turns out three times as many meals. How?
Story one: better cooks. The kitchen bought twice as much of everything, and the cooks became 50% faster at turning ingredients into meals.
Story two: cheaper vegetables. The cooks are no faster. But prices fell by a third, so the doubled budget bought three times as much of everything.
Story three: a bit of both. Prices fell a little, and the cooks got a little faster. There is not one such mix but a whole range of them, from all better cooks to all cheaper vegetables, and every one fits the same two numbers.
Move the slider. The budget and meal bars never change.
Why?From outside, with only the budget and the meal count, you cannot tell which of these kitchens you are looking at. You would need the price list, or a look inside the kitchen.
Anthropic's numbers have the same problem. We see what was spent and how fast the score rose. We do not see how the money was split between people, experiments and training new models. We do not see what a unit of computing cost. And AI help is itself bought with money, because it runs on computers the company pays for. So "money" and "AI" are not two separate piles that the speedup can be sorted into.
There is a further twist. If Anthropic spent more because AI tools made research pay better, then part of AI's effect works through the extra spending. So there are two questions, not one: what AI did with the money held fixed, and what it did counting the extra money it attracted. They have different answers, and both are worth asking.
What the public numbers support is narrow, and it cuts both ways:
The day before we finished this page, two economists, Luis Garicano and Alex Imas, published a fuller study of the same numbers. They brought two things our sections above lacked: the scores of sixteen Claude models over time, read from METR's chart, and estimates of how many technical staff Anthropic had.
Over a longer window than ours, they find what Wiebe found. Progress per dollar fell about 69%, close to the 63% above. Progress per technical employee fell about 38%. Spending per employee roughly doubled, which they read as more computing per researcher.
Then they did something our sums could not. They took a standard economist's model of research, in which progress depends on how much effort goes in and on how hard the next idea is to find. They fitted it to the six models before the bend. Then they fed in the money and staff that came after, and asked: does more effort alone explain what happened?
Their answer: not fully. In their most generous version, more spending explains about half the speedup. Measured other ways, it explains much less. The rest looks like a one-time jump in how well research works. They suggest coding agents as one possible cause, and say plainly that their figures are model comparisons, not measured effects of AI tools. Their own caveats matter here: the final score still falls inside the margin of error of their effort-only forecast, and in the most generous version the jump is not statistically clear.
So their conclusion sits near Krier's: progress sped up partly because more went in and partly from a one-time boost whose size and cause are uncertain, and ideas keep getting harder. The per-dollar chart could not show this. It took the extra data.
There is one more thing the spending numbers can illustrate, though not prove.
Research gets harder as it goes. The easy discoveries are made first. To keep progress steady, a field needs more and more researchers and equipment each year.
How much more each year? The research machine, an optional page later in this series, assumes a figure: about 1.4 times a year. Anthropic's own spending hints at a much steeper climb. Through 2025 its research spending grew somewhere between 2.8 and 5.9 times a year, depending on which months you count. Both figures include a few months from after the speedup, and spending is not the same as research effort, as section 4 showed. Garicano and Imas, in an illustration in their appendix, find that holding a pace of about 25 points a year would need research spending to roughly quadruple each year, on their assumptions.
A worked example on the optional research machine page shows how the workforce needed to hold a doubled pace can grow under its assumptions.
This staircase describes holding a faster steady pace, the second shape on the overview; the third shape is an explosion.
What survives is the difference between a step and a staircase. Speeding up once is a step. Staying that fast, while research keeps getting harder, means climbing a new, steeper step every year. Krier made a related point in his first post: a one-off jump of 1.9 times "isn't the same as sustained exponential" growth. On METR's scale, roughly a fixed number of points stands for a doubling of the length of task an AI can handle, so a steady pace in points is already exponential growth in task length. The question is whether the pace itself keeps rising. The worked example illustrates how hard holding the pace could be on its assumptions. It does not measure it.
These numbers cannot settle the puzzle on their own, and more numbers of the same kind would not settle it either. Here is what would narrow it.
This page is also a record of how far AI work goes with little steering. Here is how it was made.
What to make of it: the AI systems did nearly all the sums, the code and the checking. They also made the biggest mistakes, in both directions, and other AI systems caught them. The human asked the question, noticed when the words stopped making sense, and decided to ask for a second opinion.
Your guess
You wrote this at the start. Would you change it?
Next time someone tells you that something "per dollar" has fallen, here are four questions to ask:
Built by Ashish Kulkarni with Claude and Codex. The argument was reviewed by Astra. Claude is made by Anthropic, the company this page is about. Code and text are open. Built to be argued with.
Design craft learned from Keshav Jindal.