I have never once opened my token usage dashboard and I think that might be the closest thing I have to a principled position these days, because the people around me talk about their token counts the way a previous generation talked about lines of code and every conversation makes me a little less interested in knowing my own number. Somewhere in the last year the units of work became the work, and the whole industry went along with it because consumption is easy to measure and output is hard, so we built leaderboards for the thing that is measurable and let the thing that matters float around undocumented.

The numbers out there are so absurd they read like a parody someone drafted and then refused to publish. Meta ran an internal thing they called Claudeonomics where people competed on how many tokens they could burn through and the leaderboard totaled sixty point two trillion tokens in thirty days with one individual sitting at two hundred and eighty one billion which is a number so large it stops meaning anything and just becomes a flex. Uber apparently blew through a three point four billion dollar annual AI budget in four months because that is what happens when a budget is treated as something to use rather than something to protect. There is an engineer out there with a monthly AI bill over a hundred and fifty thousand dollars and an OpenAI employee who consumed two hundred and ten billion tokens in a single week, which at this point stops being engineering and becomes a spiritual practice that happens to bill by the thousand.

I keep thinking about how this is the exact same mistake as lines of code, which the industry spent decades embarrassing itself out of. Lines of code was a metric that measured effort without measuring outcome, so it rewarded the people who generated the most text and punished the people who wrote the delete button. Token consumption is that same metric dressed in a much more expensive suit because now the bloat has a direct cost per thousand and the incentives to game it are even worse, since a leaderboard that ranks people on consumption is a leaderboard that teaches people to consume. The scary part is that the cost is only the beginning, because when you rank engineers on tokens the person who spends three hours refactoring a hundred lines away looks lazy next to the person who fed the whole codebase to a model and pasted the result back, and that reward inversion warps the judgment of everyone watching.

Why does everyone rationalize it though, because that is the part that interests me more than the waste. Management needs a signal and reaches for the one that arrives as a number instead of the one that arrives as an argument, since a token count lands on a spreadsheet while a shipped feature has to be defended in a meeting. Founders and executives who want to sell the story that spending on AI is adoption and adoption is progress push proposals around like the one where a five hundred thousand dollar engineer should burn at least two hundred and fifty thousand in tokens a year, and the moment a budget becomes a review the behavior warps to feed the budget. And then there is us, the engineers, because when a boss tells you a number is good you learn to make the number good, and there is a particular kind of person who discovers they can run up a bill and call it curiosity and nobody argues because the metric is on their side.

The engineering skill that actually matters has never been about the volume of anything, and it is embarrassing that we keep needing to relearn this. The skill is doing the thing with the least friction and the least fuel and the most surviving code, and I have seen enough of both to know the difference. I have watched a colleague solve a problem with a tight prompt and a small context window and I have watched the same problem get solved by someone who loaded the entire repository in and let the model churn for an hour, and the first version was faster and cheaper and more correct while the second produced a diff so large nobody could review it, which is how you get the statistics about bug rates climbing fifty four percent and rework jumping eight hundred and sixty one percent. Burned tokens do not ship features. The companies that measured by output instead of intake noticed that the people with the giant bills were not the people shipping the most.

I do not think the era of these leaderboards lasts much longer because the money runs out before the enthusiasm does, and a farce like this always ends the same way, with someone adding up the column and noticing the bill did not buy anything it was supposed to buy. The wise managers will quietly delete the dashboards and start asking what shipped and what survived and what it cost, and the rest of us should stop treating our token counts like a bank balance, because consumption is not accomplishment and I say that as someone who has never even looked at his own.

Tell me what you spent last month. I am @troysk704.