The Human Token Economy

Saw a thread this week about AI agents now burning through five times more tokens than humans do for the same tasks. Someone asked, quite reasonably, what a “human token” even is. Someone else answered: it’s the moment on a Monday morning when your boss makes the dumbest request of the week, and you decide to swallow it and get on with your day instead of telling him where to shove it. That’s one human token, spent. Their quota runs out by 9:15am most days.

I laughed at that harder than I’ve laughed at anything work related in a while, mostly because it’s true and mostly because it maps so neatly onto the actual technical discussion happening underneath it, which is: nobody’s entirely sure what they’re measuring.

There’s a real argument in that thread worth sitting with. One person pointed out that a lot of this token blowout isn’t some grand scheme to bill you more, it’s engineers at the big labs operating as if token cost will collapse to zero eventually, so they’re not bothering to optimise for it now. They’ll torch two hundred grand on a test without blinking. Another person called bullshit, more or less, saying that’s still compute they could be selling and it doesn’t add up unless the real goal is just being first, cost be damned. Both of those things can be true at once. I don’t think there’s a tidy resolution here, and I’m suspicious of anyone who tells you there is.

The bit that actually landed for me, though, was the comment about token usage being a rubbish proxy for productivity. An agent can spend five hundred thousand tokens because it solved something genuinely hard, or it can spend the same amount going in circles, rereading the same three files, trying the same broken approach four different ways with slightly different phrasing. From the outside those two things look identical. What you actually want to know is cost per accepted result, how much did it burn before it produced something a person actually kept. Nobody’s measuring that properly yet, as far as I can tell, and until they do, “5x more tokens” is a headline without much underneath it.

I use agents for work most days now, DevOps stuff mostly, scripts and config and the occasional bit of glue code nobody wants to write by hand. Some days it’s genuinely great, feels like having a very fast, very literal junior who never gets tired. Other days I watch it thrash on something trivial for ten minutes, burning tokens like it’s got somewhere better to be, and I end up doing it myself in ninety seconds out of sheer impatience. Both experiences are real. I don’t think one cancels the other out.

What nags at me is the electricity underneath all this. Every one of those wasted retries is a data centre somewhere pulling power, and we’re nowhere near being honest as a society about what that’s going to cost us, in dollars or in emissions. I like the technology. I’m properly fascinated by it, some days. But “engineers acting as if cost is free” is a fine strategy right up until someone has to pay the actual power bill, and that someone is eventually all of us.

The comment that stuck with me longest was the one about the incentives: giving OpenAI or Anthropic five times the money for shoving a non-deterministic system into a job that used to be deterministic. That’s the tension nobody in the thread quite resolved, and I don’t think I can either. Sometimes the extra spend buys you something real. Sometimes it’s just the machine talking to itself, expensively, while you wait.