Skip to content
Fritz Gerald ZephirinMenu

Social

The Cost of Intelligence Is Falling. The Cost of Judgment Is Not.

4 min read

In November 2023, OpenAI cut the price of its best model by roughly two-thirds. Six months later, GPT-4o arrived at half the price of its predecessor. In December 2024, a Chinese lab trained a frontier-class model for a disclosed $5.6 million of GPU time. Whatever else is true about AI, one fact is now beyond argument: the price of machine cognition falls fast, and it falls on a schedule.

Here is what hasn’t fallen: the cost of deciding what to ask for. The cost of knowing what “done” looks like. The cost of standing behind an answer when it’s wrong. I run customer support operations, and I watch this asymmetry every day. The model’s draft reply costs a fraction of a cent. Deciding whether that reply actually resolved the customer’s problem — and being accountable when it didn’t — costs what it always cost.

Economics has a name for this pattern. When the price of an input collapses, its complements appreciate. Bandwidth collapsed in the nineties, and the value moved to the companies that decided what was worth your attention — the carriers got commoditized; Google got the profit pool. The spreadsheet eliminated calculation as a profession and multiplied analysis as one: once arithmetic was free, knowing what to model became the job. Cheap intelligence is doing the same thing to knowledge work, and the appreciating complement is judgment.

I want to be precise about that word, because “judgment” can hide a lot of hand-waving. I mean three specific things. Preference: knowing what you actually want, which no model can supply because it is not a fact about the world. Standards: knowing what counts as done, which in operations is the difference between a deflected ticket and a resolved one. Accountability: being the one who signs. When Google says AI now writes more than a quarter of its new code, the load-bearing clause is the one people skip: “then reviewed and accepted by engineers.” The generation became cheap. The acceptance did not.

The strongest objection is obvious: models will do judgment too. Take it seriously — model evaluation skills improve every quarter, and the length of tasks AI can complete keeps doubling. But notice what improves and what doesn’t. A model can increasingly tell you whether code is correct. It cannot want the product to exist, set the bar for what ships, or absorb the liability when the bar was wrong. Those aren’t capability gaps that scale away; they’re features of being the principal rather than the agent. Even a model that judges better than you doesn’t answer for the judgment. Someone still signs.

This is also, I think, the most honest reading of the era’s most embarrassing statistics. MIT found that about 95 percent of enterprise AI pilots produced no measurable P&L impact. An NBER survey of six thousand executives found roughly 90 percent reporting no productivity effect at all. The models were not the constraint in those companies. Capability was abundant and judgment was scarce: nobody decided what the tool was for, what the baseline was, what “working” would mean, or who owned the outcome. The intelligence was purchased. The judgment was assumed.

The operating implications are concrete. If you sell knowledge work, price the judgment, not the output — output is the part collapsing toward zero. If you run a team, expect the shape of work to invert: fewer people producing, more people specifying and verifying, and the verification layer becoming where quality actually lives. If you’re hiring, notice that the market now overpays for credentials that signal production capacity and underpays for demonstrated judgment — that mispricing is an opportunity. And if you’re early in a career, the uncomfortable news is that judgment was traditionally built by doing the cheap work that machines now do. That problem is large enough that it gets its own essay.

A falling price is not a small thing. It is the kind of thing economies reorganize around — and this particular price drop touches the input that white-collar work is made of. But the reorganization will not reward the companies that consume the most cheap intelligence. It will reward the ones that figure out, faster than their competitors, where judgment actually lives in their business — and concentrate it there.

The cost of intelligence is falling. The cost of judgment is not. Most strategy questions of the next decade are that sentence, applied.

Sources & further reading