Thinking Got Cheap. Work Hasn't Caught Up.

Since 2023, the price of getting a GPT-4-level answer out of a machine has fallen at least tenfold a year by most measurements, and on some benchmarks much faster.

That is the part of the AI story everyone sees, and it's the least useful part to stare at. A cheaper engine doesn't tell you much about who wins, which jobs change, or what you should do on Monday. Three other things do, and they move at very different speeds. Most of the confusion about AI, the hype and the disappointment, comes from mixing them up.

Scroll sideways to see the whole figure

Clock one: the engine keeps getting cheaper, and the bill doesn't fall

Start with what you actually pay for. A language model reads and writes in tokens, chunks of text a bit shorter than a word, and every token costs a slice of computation on a very expensive chip. Providers price per million tokens. When people say "AI got cheaper", they mean that the price of a token at a fixed level of quality keeps falling: the endpoints of Epoch AI's price data give about 10 to 17 times per year, depending on which level of capability you hold constant, and Epoch's own fitted trends are often steeper (Where the AI Money Goes).

So why are companies spending more on AI, not less?

Because nobody buys the old quality at the new price. They buy the new quality, and the new models think out loud before answering. A reasoning model can write thousands of hidden tokens to produce a three-line reply. Agents loop, read files, call tools and try again. Ramp, which sells spend-management software and sees what its business customers pay for AI, reported their token volumes growing about elevenfold between January 2025 and April 2026 while spending grew about sixfold. The price per token fell. The number of tokens bought rose faster.

Cheaper thinking doesn't shrink the bill. It moves it to the tasks nobody could afford to automate before.

This is an old pattern in economics wearing new clothes. In 1865 William Stanley Jevons argued that as coal-burning engines became more efficient, Britain burned more coal, not less, because suddenly it paid to use engines everywhere. Cheap thinking is following the same path. The useful question has moved from "can we afford to use AI here?" to "what would we do if a competent first draft of anything cost a few cents?" (What Becomes Possible Now?).

Clock two: it gets smarter unevenly, and you can predict where

The second clock is capability, and it does not tick at the same rate everywhere.

Two years ago the standard complaint was that models couldn't count, couldn't do arithmetic, invented citations. Today the best ones solve olympiad problems, and coding agents fix real bugs in real codebases. METR, a research group that measures how long a task (in skilled-human time) an AI agent can complete half the time, has published data in which that horizon doubles roughly every four months since 2023. Meanwhile the same models still write a mediocre strategy memo, and still occasionally state a wrong fact with complete confidence.

That unevenness has a cause, and the cause is predictable: models improve fastest where an answer can be checked cheaply and automatically. Training increasingly works by letting a model try a problem many times and rewarding the attempts that pass a check: the unit tests run, the proof verifies, the number matches. Where a computer can grade the answer, labs can run that loop millions of times. Where grading needs an expert's afternoon (is this strategy sound? is this contract clause wise?), they can't, and progress is slower (Why Progress Arrives Unevenly).

That gives you a practical forecasting tool, better than any leaderboard. For any piece of your own work, ask: could a computer tell whether the answer is right? Reconciling two spreadsheets, extracting fields from invoices, translating, writing code against a test suite, checking a document against a rulebook: yes, or nearly. Expect fast progress there, measured in months. Choosing which market to enter, judging a candidate, deciding what a customer really meant: the check is expensive or contested. Expect slower progress, and expect the model to sound just as confident there as anywhere else.

The industrial world shows the same pattern outside software. Nissan reported in November 2025 that a model from the AI company Monolith, trained on its engineering data, recommended torque ranges for bolt joints in vehicle chassis and ranked which further tests were worth running, cutting physical testing in that pilot by 17 percent. Bolt joints come with a hard, physical check at the end, the test rig, and decades of recorded results. That is the kind of place the checking argument says AI should bite early.

Clock three: the factory hasn't been rebuilt

The third clock is the slowest, and it's the one that decides who actually benefits.

By 1900, American factories had been able to buy electric motors for nearly two decades, and motors still supplied less than five percent of the mechanical power driving their machines. Manufacturing productivity didn't surge until the 1920s. The reason wasn't the motors. The first factories that electrified simply replaced the big steam engine with a big electric motor and kept everything else: the central drive shaft, the belts, the building laid out around power rather than around the work. The gains came only when a new generation of factories put a small motor on every machine and reorganized the floor around the flow of materials.

We're in the "one big motor, same old belts" phase of AI.

Most companies that have adopted AI started by installing the engine. They gave people a chat window, bought seats, ran a pilot. Then they measured, and the numbers were confusing. A careful 2025 trial by METR found experienced open-source developers took 19 percent longer on their own projects with AI tools, while believing they'd been faster. In a large customer-support study, agents with an AI assistant resolved about 15 percent more issues per hour on average, and the lowest-skilled fifth gained about 36 percent. Similar technology, opposite results.

Part of the difference is who used it: the support gains went mostly to novices, while the developers were experts in code they knew intimately. The rest is what happened around the tool. In the support study, the assistant sat inside the flow of the work and carried the best agents' practices to everyone else. In the developer trial, the AI sped up writing code, while the slow parts of the job (understanding the codebase, reviewing, integrating) stayed exactly where they were, and checking the AI's work took time too. Speed up one step of a process and the time simply piles up in front of the next bottleneck (Where the Saved Time Goes).

Installing AI is the easy part. The productivity is in rebuilding the work around it, and that has barely started.

That's why "AI isn't delivering" and "AI is transforming everything" can both be said, sincerely, about the same company in the same quarter. The engine works. The factory floor hasn't moved.

When thinking is cheap, something else becomes expensive

Here's where it gets interesting for anyone running a business, or planning a career.

When one input to production becomes nearly free, value doesn't disappear. It moves to whatever must be combined with that input and didn't get cheaper. When drafting, analysis and code become abundant, four things become the constraint:

Scroll sideways to see the whole figure

Checking. If a model produces ten proposals in the time a person produced one, someone still has to know which one is right, and answer for it. The person who can verify quickly and confidently in a domain becomes more valuable, not less. That is also why a model that's "95 percent accurate" can be worthless in one process and transformative in another: it depends entirely on how cheap the remaining five percent is to catch.

Context. A model only knows what reaches it. The exceptions, the reasons behind a rule, the customer who always pays late but always pays, the decision that was made in 2021 and why: much of what makes a company work lives in people's heads. Organizations that write down the why along with the what give every future model a better place to work. That was good advice for training new employees. It's becoming the main thing that separates an AI that's useful from one that's merely articulate (Write for the Model That Doesn't Exist Yet).

Access and trust. The model that can read your ERP, act in your CRM and be trusted to send an email on your behalf is worth far more than a smarter model that can do none of that. Permissions, integrations, security reviews and track records take months to earn, and no model release shortcuts them (A Product That Improves With the Model).

Distribution. When everyone rents roughly the same models, the better model gives you a lead that lasts until the next release. Owning the moment where the work happens lasts longer.

That last point has already rewritten the industry story many people believed in 2023. Google invented the transformer and then hesitated, for good reasons that looked like bad ones, while OpenAI ran ahead. Plenty of people concluded that the incumbent was finished. Instead, Google put AI summaries on top of the search results everyone already used, and by July 2025 it reported more than two billion monthly users seeing them. The answers didn't have to beat ChatGPT's; nobody had to go anywhere to get them. Microsoft did the same through Office and GitHub. The disruptors who've thrived, like ChatGPT itself and the coding tool Cursor, did it by creating a new place to work (Distribution Beats the Better Model).

Incumbents don't always win. The contest has turned less on the model than on who owns the place where the work happens, and who has the context, the access and the trust once they're there (Where the Moat Moves).

So what do you actually do?

If you lead a company, a team, or just your own work, the three clocks suggest a posture that's neither hype nor waiting. It is pragmatic, and slightly contrarian on a few points.

Adopt now, optimize later. Many of the tools you choose today will be obsolete in a year. That argues against over-investing in any single one, and against waiting. Start with what already exists, use it on real work, and learn where it helps and where it lies. The learning compounds even when the tool doesn't survive.

Stay a step behind the frontier for what matters, and play at the edge for what doesn't. The newest technique of the week is usually unproven and often already absorbed into the next model release. For your core processes, use methods that have held up. For experiments, go wild. Most of the exhausting "keep up with AI" pressure comes from mixing the two.

Keep switching cheap. Don't marry a provider. Keep your own record of how the work runs, your own test cases, your own data, so that a better or cheaper model next year is a configuration change, not a migration project (Choosing a Model Provider Without Marrying It).

Write down the why. The single most future-proof investment most companies can make is documentation that captures intent, exceptions and reasons. Every model generation will be able to use it better than the last.

Measure outcomes, not output. "We produce 60 percent more proposals" is output. Whether you win more deals, at what margin, and what your people stopped doing to produce them is the actual result. The gap between the two is where most AI business cases quietly die (Output, Outcome and Value).

Invest in people learning to drive, not memorizing the engine. Most people don't need to understand attention mechanisms. They need to know what the tool is good at, how it fails, and how to check it, which comes from using it on their own work, imperfectly, early (Learn to Drive Before You Study the Engine). Training budgets with no immediate return look like waste right up until they're the reason your company adapts faster than the one next door.

The steam engine, again

In 2023 the safe-sounding position was to wait for the hype to settle. It was the wrong call then and it's a worse one now. AI looks much less like a tech bubble than like electricity or the steam engine: a general-purpose shift whose gains arrive late, unevenly, and mostly to the people who reorganize around it.

But the electricity story carries a second warning, easier to miss. Electricity didn't reward the factories that bought motors first. It rewarded the ones that rebuilt the floor. Roughly four decades separated Edison's first power stations from the productivity surge.

The engine is already cheap, and it'll be cheaper next year. The real question for anyone reading this is which part of their own factory they're going to rebuild, and how much faster than last time.

Go deeper

Each section above opens into pieces that go further:

• Costs and prices: Where the AI Money Goes · Why AI Products Have Thinner Margins · What Thinking Buys
• Where capability grows: Why Progress Arrives Unevenly · Confidently Wrong · Leaderboards Measure the Test
• Rebuilding the work: Where the Saved Time Goes · What Becomes Possible Now? · From a Good Demo to a Working Department · A Job Is a Bundle of Tasks
• What gets scarce: Write for the Model That Doesn't Exist Yet · Data Access Is the Real Integration Project · Human Review Where It Matters
• Who wins: Distribution Beats the Better Model · Where the Moat Moves · Why Labs Build Products · Who Receives the Gain?

Check yourself

Why doesn't cheaper thinking shrink the bill?

Nobody buys the old quality at the new price. New models think out loud and agents loop, so the number of tokens bought rises faster than the price per token falls.

Why do models improve faster at code with tests than at strategy memos?

Training works by rewarding attempts that pass a check. Where a computer can grade the answer, labs can run that loop millions of times. A strategy memo needs an expert's afternoon to grade.

What happens to a company that buys AI seats but leaves its work unchanged?

It is in the "one big motor, same old belts" phase. The engine works, but the productivity comes from rebuilding the floor around it, as with electricity.