Why we’re mispricing AI intelligence, and why “good enough” is the real revolution
The week the bill arrived
What finally pushed me to write this wasn’t a new model. It was a week in the stock market.
In late October, Anthropic, a company founded only in 2021, filed to go public. Investors had been talking about a valuation of around $2 trillion. Within a week, after the company shared its financials, that talk had dropped to about $1.5 trillion Source. Then its biggest partners reported earnings, and the mood shifted. Microsoft, Google and Amazon all reported infrastructure spending below analyst estimates. Hyperscaler cash flows were being squeezed, Microsoft turned to the bond market to help fund its buildout, and the Nasdaq had its first three-week losing streak since March Source. The debt-funded “neoclouds” had it worst. CoreWeave fell 30% between the end of September and mid-October, while the S&P 500 rose slightly Source.
Let me be fair about this. The demand is real. Cloud revenue is booming, and Anthropic’s prospectus showed roughly 1,300% annual sales growth and an implied 85% gross margin. But a Wharton professor asked the question I keep coming back to: “It’s a great technology, seems to do things that we never would have imagined, but how does that then translate into a viable business?” Source
When the richest companies on earth have to borrow to keep up with the intelligence race, it’s fair to ask whether we’re putting the right price on intelligence at all. I don’t think we are. And I think the mistake is a very old one.
The moment things changed
I’ve been coding with AI agents for over a year, and I can tell you when it all changed: late November 2025, when Anthropic released Claude Opus 4.5. OpenAI released GPT-5.1-Codex-Max around the same time.
What made that generation special wasn’t that the models suddenly got smarter in some abstract sense. They got durable. Both companies leaned on a technique called compaction: the model summarises its own earlier work so it can keep going across a long task instead of hitting the limit of its context window Source. Anthropic put it simply: “long conversations no longer hit a wall.” They also cut Opus pricing to $5 per million input tokens and $25 per million output tokens Source. Independent evaluators estimated these models could complete, about half the time, software tasks that take a skilled human several hours. For the first time, you could hand an AI something that felt like a real job, walk away, and come back to finished work.
A month later, Andrej Karpathy, one of the most respected engineers in AI, posted: “I’ve never felt this much behind as a programmer.” Source Since then, the story has spread from code into mathematics, physics and research, and a familiar idea has taken hold: intelligence is the bottleneck, so whoever builds the most of it wins everything.
One detail from the Opus 4.5 launch has stayed with me. Anthropic said the model scored higher than any human candidate ever on its notoriously hard engineering take-home exam. In the same breath, the company noted that the exam “doesn’t test for other crucial skills candidates may possess, like collaboration, communication, or the instincts that develop over years.” Source
The people who built the smartest coder in the world drew a line between intelligence and everything else that makes someone valuable. The market hasn’t.
We have always overrated intelligence
Think of the most brilliant person you’ve ever met. Are they the most successful person you know? Usually not. The very top academics are often a little out of step with the world. They struggle to communicate, clash with colleagues, or get lost in problems no one is paying them to solve. In most corporate settings, the smartest person in the room rarely runs the room.
The data supports this. A study of more than 7,000 Americans by Jay Zagorsky at Ohio State found that the income gap between an average IQ (100) and the top 2% (130) was only $6,000 to $18,500 a year. On wealth, he found no meaningful link to IQ at all. His summary: “Those with low intelligence should not believe they are handicapped, and those with high intelligence should not believe they have an advantage.” Source
A Swedish study went further, linking full national income records to military cognitive tests taken by 59,000 men. Ability rises with pay up to about €60,000 a year. Then the curve goes flat. The top 1% of earners, who make twice as much as the people just below them, actually scored slightly lower on the tests Source.
Malcolm Gladwell made a similar case in Outliers. He argued that beyond an IQ of about 120, extra intelligence makes little measurable difference to success Source. He contrasts two geniuses. Chris Langan has an IQ of around 195, grew up poor, and spent much of his life in ordinary jobs. Robert Oppenheimer, as a Cambridge student, tried to poison his tutor, got away with probation, and went on to lead the Manhattan Project Source. The difference wasn’t brainpower. It was an upbringing that taught Oppenheimer how to deal with authority Source.
Then someone always brings up Elon Musk: the richest man in the world, and obviously very smart. But no one has ever called him smooth. Alex Gibney’s new four-hour documentary Musk doesn’t even present him mainly as a genius inventor. It portrays him as a kind of Gilded Age confidence man whose real gift is “pumping stocks and dazzling investors and the public alike with visions of a high-tech utopian future.” Source Whatever you think of that verdict, if the best example of intelligence paying off is credited with storytelling and capital-raising, the example works against the argument.
The diminishing returns have arrived in AI
Here’s my confession as a working developer: my main coding agent isn’t a frontier model.
I use Qwen 3.7 Plus, which costs about $0.32 per million input tokens and $1.28 per million output tokens Source. I’m now testing Alibaba’s newer Flash model, which comes in at roughly $0.16 and $0.47, about 12 times cheaper than Qwen 3.8 Max Source. On many of my tasks it already beats Plus. The top tier, Qwen 3.8 Max Prime, is listed at $4 and $12 per million tokens Source.
Does Max write better code? On a single attempt, probably. But real software isn’t written in one attempt. It’s built in loops: write, test, fix, repeat. In a loop, a cheaper model that’s “smart enough” catches up by simply trying again, and the price gap grows with every token. One independent review found the small Flash model matched a pricier competitor with perfect scores on math and multi-step agent tasks, and fell behind mainly on visual polish like 3D rendering Source. For most of what I build, the extra intelligence doesn’t pay for itself.
That’s the Swedish curve again. Intelligence pays off until you have enough for the job. After that, other things decide the outcome.
Economist Tyler Cowen describes the wider version. When you suddenly get much more intelligence, “the other constraints in your system become a lot more binding, that the marginal importance of those goes up, and the marginal value of more and more IQ or intelligence goes down.” Source A super-intelligent model doesn’t speed up a clinical trial, approve a building permit, or win over a sceptical committee.
COVID already showed us which jobs AI could take
Here’s something we seem to have forgotten.
Before 2020, we were told business ran on handshakes, education needed classrooms, and deals needed flights. Then COVID ran a three-year experiment on all of us. Classes moved to Zoom. Business travel stopped. Consulting, law, finance and software carried on with barely a pause. Economists Jonathan Dingel and Brent Neiman estimated that 37% of US jobs could be done entirely from home Source.
We should have seen the next step coming. Once a job runs entirely through a screen, it makes little difference whether the person on the other side is in a spare bedroom or is an AI agent. COVID turned about a third of the economy into text, video and files, which is exactly what language models work with. It’s no accident that coding, consulting and analysis are where AI has landed hardest.
Now look at the other two-thirds. The baker, the barista, the plumber, the care worker, the person flipping burgers: none of them went home during COVID, because their work happens to physical things in the real world. Today, according to the US Census Bureau, only about one in five American businesses uses AI in any business function. In the Information sector it’s about 40%. In retail it’s about 14% Source.
Roboticist Hans Moravec spotted the irony back in 1988. It is comparatively easy, he observed, to make computers perform like adults on intelligence tests, and very hard to give them the perception and mobility of a one-year-old. We’ve built machines that pass graduate exams. A robot that reliably makes your coffee, restocks a shelf and calms a confused customer is still mostly a demo. The bottleneck in the physical economy was never intelligence. It’s hands, legs, sensors, liability, regulation and margins.
Pricing AI as if it will transform the whole economy at the speed it transformed the Zoom-able third is the same mistake as believing the smartest person in the room will end up running it.
Why is everyone fighting the same war?
I’m not saying frontier models are useless. Drug discovery, protein design, unsolved mathematics: there you may really need the very best. But that’s a niche, and a niche doesn’t need half a dozen companies each spending hundreds of billions racing to the same summit.
The Manhattan Project is a useful comparison, and Oppenheimer made it vivid for a new generation. The race for the bomb involved a handful of national programmes. Not every industrial company in America tried to build its own. Some problems call for a few concentrated efforts at the frontier, with the results then shared. Treating frontier capability as the main event for the entire economy is like putting every student through a PhD because a few will win Nobel prizes.
That brings me to the Ivory Tower. Andrew Rossi’s 2014 documentary of that name asked why American higher education had become so expensive, and whether families were overpaying for a promise that brainpower is the ticket upward Source. Economic historian Gregory Clark gave an uncomfortable answer in The Son Also Rises. Tracing surnames across centuries, he argues that social mobility is lower than standard estimates suggest, is about the same across societies, and resists the policies meant to change it Source. Economists dispute his numbers Source. But the core intuition is hard to shake: handing people more intelligence, whether through a diploma or an API key, doesn’t automatically change where they end up. Family, networks, capital and luck still do much of the sorting.
To be fair to the other side
The strongest counterargument goes like this: today’s cheap models are good because of the frontier race. The Flash model I love stands on the shoulders of last year’s giants, and that progress had to be paid for by someone. That’s true. But it argues for some frontier research, not six labs and a debt-funded data-centre boom all chasing the same benchmark.
The second counterargument is that people really are paying for the top tier, and Anthropic’s revenue proves it. Also true. But look at what made Opus 4.5 special: stamina, long context, staying coherent over hours of work. That’s less about raw IQ than reliability. The market is paying for something real, and it may be calling that thing by the wrong name.
The paperless office, again
In the 1990s we were promised the paperless office. What came first was the opposite. Email made us print more, because now we printed our emails too. Paper use in the US peaked around the turn of the millennium and then fell steadily, eventually by about a third. Researchers credit the decline mostly to two things: a second invention, the smartphone, and simple time for habits to change Source.
That’s how I think about AI. I don’t doubt it will reshape the world. I’ve watched it reshape my own job in a year. The internet was a real revolution too, and the stock prices of 2000 were still wrong. A technology can be right while its timing and its price are wrong.
Today’s valuations assume intelligence is the scarce ingredient, and that the whole economy is waiting for more of it. My experience as a developer, decades of research on human success, and three years of COVID all point the other way. Past “good enough,” intelligence stops being the constraint. What holds the world back is bodies, institutions, trust and time.
Intelligence is necessary, but it isn’t enough. And right now, it might be the most overpriced thing in the room.
Leave a Comment