Remember a year back when Sam Altman said something like “Stop being nice to AI, it wastes money?” That comment set me off. At the time, I was using AI as a thought partner. Being polite, being curious, having a conversation—that was the experience. Who cares if it cost a few extra tokens?
What a difference a year makes. Now I’m deep into agentic coding at work, and I get exactly what he meant. Those “all you can eat” subscriptions people use at home? They don’t exist for businesses. Every prompt, every response costs money. And once teams start adopting the tech, the costs add up fast.
So how do you prevent Bob from Ralph looping all weekend and using all the tokens? Unless you're Google, you impose spending limits. Per person. Blow through your tokens? You're done for the month. Or you go beg DevOps for more.
But once you’ve felt the speed, there’s no going back. Working this way feels like flying down the Autobahn at 500 miles per hour. Features, bugs, docs -- those dreaded unit tests? That stuff that used to take weeks? It gets done in ... days. You feel like a superhero. You feel unstoppable. Until you hit your usage limit.
Coming off a 50 PR-a-day high? It hurts. Bad.
So you have to learn to be conservative, to be token responsible. You walk into every session with rules:
“Be terse.”
“No preamble.”
“No recap.”
And that's where the frustration starts. Because the way to learn new tech is not by being restrained. It's by being curious. You need to try things out, push yourself, ask questions. Make mistakes. That's when the Aha! moment happens. That's when you start seeing the pieces come together. But wandering costs money and every decision has implications:
Long, vibey sessions? Gone.
Politeness? Optional.
Better models? Use sparingly.
You start watching your usage meters. You start thinking about tokens while solving problems. Sometimes you think about tokens more than the problem itself.
Can I risk asking Opus for help? Is Sonnet going to put me in a hedging loop? How much do I have left to spend?
It’s a low-grade tax on your brain. And it doesn’t just change how you work, it changes how the team works. Because if tokens are a scarce and valuable commodity, people will naturally protect them. They're not bad teammates. They're just trying to get their own work done before the meter runs out.
And that’s the part that I didn't expect: the cost pressure. I guess I was just naive. But now I'm starting to feel like AI is becoming less of a collaborator and more like a tool I'm forced to manage. It adds unexpected anxiety to the equation.
There's no good fix for this problem. These systems are expensive to run, subsidies are drying up, and companies are feeling compelled to get into the race.
Which means this tension isn't going away. If anything, it's going to get worse. And session and token management is going to become real friction in all our lives.
So now I basically work in two modes. At work I have to be efficient and focused. At home, I get to be curious again. I wander and experiment. I waste tokens.
And I remember why this felt like so much fun in the first place.