When AI Stops Being Cheap

In June 2026, Uber capped how much its employees could spend on AI after the company blew through its budget in four months. Around the same time, Salesforce disclosed it was spending three hundred million dollars on Anthropic tokens while freezing engineering hires, and Microsoft moved to cancel Claude Code licenses across the division that builds Windows, Office, and Teams, steering thousands of engineers toward its own GitHub Copilot as the bills for all that usage mounted. In each case the tools worked and got used; what broke was the budget, as real usage ran into costs nobody had planned for.

For two years everyone assumed AI would only get cheaper to run, and for a while it did. Tomasz Tunguz of Theory Ventures recently charted three years of model pricing and found that reversing: Google's Gemini has tripled year over year, OpenAI's flagship is climbing again after a subsidized stretch, and the most capable models now charge around five dollars for every million tokens they read and twenty-five to thirty for every million they write. (A token is the small unit of text an AI reads and writes, and what it is billed for.) His read is that vendors cut prices when cash is plentiful and market share matters, and raise them when cash is tight and margins matter. With spending on data centers at record highs, all three of the major players are now in the second mode.

None of this is a surprise. The cheap years had the feel of a honeymoon - the providers losing billions to win the market, the way ride-sharing once kept a cross-town trip at five dollars. Land-grabs like that don't last, and the bill is starting to reflect what the service actually costs.

The list price is only half of it. The number of tokens a model burns to do the same job is also rising, and the sharpest measurement of that comes from the team at Bigspin, who built a consumer price index for AI output, the same idea as the cost-of-living index, but for AI work. Across 5,851 real coding sessions this spring, the buying power of a token fell to about twenty-three cents on the dollar. What ten thousand tokens used to deliver, around a third of a finished code change and six hundred lines of code, by mid-April bought a sixth of a code change and ninety lines. The model finished about the same work per session; it just spent far more tokens getting there, with the share it spent thinking to itself, rather than producing output, rising from one percent to eighteen. Their study covers coding tools and they flag its limits, but the pattern is general: the real cost of an answer is climbing on two fronts at once, the price of each token and the number of tokens an answer takes.

The metric is shifting from tokens to results

The model makers are already marketing this way. When Microsoft launched its latest coding model, it printed the model's average token use right in the announcement, showing that it beat a rival's benchmark score using about sixty percent fewer tokens. An independent analysis ranked two leading models within a point of each other on overall quality, while one cost forty percent more to run. Same answer, forty percent more expensive.

The makers say so themselves. Amid the emerging price war, an Anthropic spokesman told the Wall Street Journal that companies now judge models on price per task, what it costs to finish a job start to finish, rather than price per token. You end up paying for the result, and for producing it the thousandth time someone needs it. That is where how a system is built starts to matter more than which model it uses.

Where the money actually goes

Most systems that answer questions about documents read the document at the moment the question is asked. One approach feeds the whole document to the model and asks; another finds the relevant passages first and feeds only those. Both are reasonable, and the second is the common default. Both also redo the same work on every question, re-reading the text and working out the same cross-references and definitions each time. For a team asking many questions of the same documents, that is the same work, paid for again and again. It is also the least reliable way to do it, since models get worse as the input gets longer even when the right passage is right there, and Stanford's RegLab has shown how often legal answers cite a source that does not actually support them.

The alternative is to do that work once, when a document first comes in, and save the result as organized data that later questions read from instead of working it out again. The hard, error-prone step happens up front, on a document that is not going to change, where the result can be checked once rather than trusted blind on every question.

Where the gap opens

How much you save comes down to two things: how often you ask, and across how many documents.

Ask one question and re-reading actually wins, because organizing the document up front costs more than simply reading it once. But organizing is a one-time cost, and every question after it is cheap: each one looks up only the part it needs instead of reading the whole thing again. Each of those lookups runs roughly ten times cheaper than a full re-read. The savings pile up fast. Somewhere around the fifth question they cover the up-front cost, and by the hundredth, re-reading the entire document every time looks absurd.

The bigger gap shows up across many documents. Ask which of ten thousand contracts renew next quarter under a given liability cap. Answering that by reading all ten thousand is wildly expensive, and more than any single model can take in at once. Against the organized version it is one lookup. That is less about cost than about possibility: the difference between a question you can ask any time and one you would never bother to ask at all.

One more saving sits on top. Not every step needs the most powerful model - the simpler parts run fine on smaller, cheaper ones, and the top model is saved for what genuinely needs it. That is much of what the price war is about, and it stacks with doing the work once: a cheaper model that no longer re-reads the same contract every time is cheaper still.

What this means if you are buying

The models will keep improving and keep fighting over price, and you don't have to pick the winner. What matters more is the system underneath: whether it keeps paying to re-read your documents, or does the work once and reuses it. As running AI gets more expensive, that gap only widens.

Through all of it, the value sits in your own organized record more than in any model on top of it. That record is what we build, but the point stands without us: treating AI as free to run is ending, and the systems that last are the ones that stop paying to read the same page twice.