Skip to content
All guidesCost & efficiency

Resisting the urge to token max

Burning through your plan can feel like getting loads done. More often you end up with a big bill and nothing you would actually put your name to. Here is the case for slowing down, staying in the loop, and spending tokens only where they earn their keep.

5 min read

If you use AI tools with a usage limit or a bill attached, you've probably had this session. You set something big running, you watch the tokens fly, and by the end you've chewed through a chunk of your plan and you're not really sure what you got for it. It felt productive. The meter was moving the whole time. But if someone asked you to show them what you'd made, you'd struggle to point at anything you would actually ship.

What token maxing is

Token maxing is going as fast and as hard as you can and treating the amount you've spent as the score. You give the AI a big, vague task, you let it loop, and you feel good because you're clearly getting your money's worth out of the plan. The purest version is typing something like "/loop make me money, make no mistakes" and walking off. It's vibe coding with the throttle stuck open. The catch is that spending isn't the same as progress, and by the time you wander back you're so far out of the loop that you can't even tell whether what it did is any good.

Why it feels like progress when it isn't

Speed is convincing. When the AI is moving quickly and the counter is climbing, it looks like a lot is happening, and it's easy to feel satisfied that you've squeezed real value out of your plan. But you can burn a whole month of usage and have nothing to show for it. Worse, because you weren't watching, you don't really understand what got built, so you can't fix it, extend it or trust it. You've paid for output that you now have to reverse engineer before it's any use to you.

Who this catches, and why

This is something we see all the time, and it's one of the most common beginner level mistakes there is. It's no dig at anyone. It tends to catch people at the beginner and intermediate stage, because the way these tools get talked about makes it sound like you can hand over something enormous and vague and let it run. And you can, eventually. Companies like Anthropic can build loops and workflows that turn out high level software almost on their own, inside a tight scope and a clear specification. They can do that because they're experts who have already done the work by hand, over and over, and know exactly what good looks like. Typing "/loop make me money" is not that. It's the same tool with none of the groundwork underneath it.

Slow down and stay in the loop

The better instinct is almost the opposite of maxing. Go slower. Get clear on what you actually want before you ask for it. Spend fewer tokens, not more, and spend them on purpose. Do the job yourself a few times, with the AI, while you watch every step. Learn how it really behaves and where it falls over. Once you've ironed a process out by hand, that's the thing worth automating, because now you know what the loop is doing when it runs. You're pointing it at a real outcome you understand, instead of setting a vague task going and hoping.

That's also when looping actually pays off. A loop you've proven is worth running on its own, because you've been through it and you can steer it at what you want. A loop you spun up from a vague sentence is just spending you aren't watching.

Spend the expensive stuff where it earns

Not every job needs your most powerful model. Save the top tier, the Fable class of model, for your hardest and newest work. Using it to summarise your inbox or answer something you've barely thought about is money spent on nothing a cheaper option wouldn't have handled just as well.

A few habits keep the rest of the spend honest.

  • Compact your sessions. A long conversation carries its whole history with it, and you pay for all of that again on every message. Clearing it down when you move to a new task keeps the cost sensible and the AI sharper.
  • Know what quietly bloats your context. Some tools you plug in, Playwright is the usual culprit, pull a huge amount of information into the AI's window every time they run. That's your tokens, spent on noise, before you've even asked for anything.
  • Reach for the powerful model last, not first. Try the cheaper path, and step up only when the task genuinely earns it.

If it is deterministic, it is a script

This is the one that saves the most over time. A lot of what people throw AI at isn't a thinking task at all. It's a fixed, repeatable job that gives the same answer every time. Renaming a batch of files, moving data from one place to another, checking a field has been filled in. You don't need a language model for that. You need ten lines of Python. Every time you turn a deterministic job into a small tool, you stop paying per run for something a script does for free, and does the same way every time.

Do that across your whole setup and the token bill comes down on its own, because the AI isn't being spent on the boring, certain jobs any more. It's saved for the work that's genuinely iterative and new, which is the only work worth paying a model to do.

What this is actually worth

All of this comes back to money, and the saving is real. On a fixed plan, being deliberate means you stop running dry halfway through the month for no return. If you're paying an API bill per token, which is where the cost can really run away from you, this is the difference between a sensible monthly figure and a nasty surprise at the end of it. Slowing down sounds like the frugal, cautious option. Really it's just the version where you understand what you're building and only pay for the parts that matter.

How we think about it at Pulsar

This is how we run AI for the businesses we work with. We're careful about where the tokens go, we build small deterministic tools for the certain jobs, and we keep a person in the loop on the work that matters, so the spend lands on the things that genuinely move the business. We've made these mistakes ourselves, so the people and teams we train and consult for don't have to. The whole aim is simple: the value you get out of the AI genuinely outweighs what you spend to get it, so there's real business value in it, not just a big bill and a pile of output nobody trusts. If your team is getting through a lot of AI and you're not sure the spend is buying you anything, that's exactly the sort of thing we sort out.