Stop Stressing About AI Token Costs (And What to Actually Pay Attention To)
At a Lindy community session I run monthly, someone asked a question I hear pretty often.
They were setting up an AI agent with a detailed context profile — basically a document that tells the agent everything about them, their preferences, their working style. Around 20 pages. And they wanted to know: wouldn’t loading all that into an agent cost too many tokens?
Fair question. Cost is real. But it’s the wrong thing to be optimizing for right now.
Here’s the analogy I used.
Remember the T-Mobile “Top Five” Plan?
For those who were texting in the mid-2000s, you had to pick five people. Five contacts you could text for free. Everyone else costs money. You thought carefully about that list. Family. Best friend. Maybe someone you were dating.
And you rationed your messages to everyone who didn’t make the cut.
Now you text anyone without thinking. iMessage, Android, doesn’t matter. You don’t know what texts cost per message. You probably never think about it at all.
That’s where AI compute is heading.
Right now, we’re paying early-adopter prices. Just like we paid per-text in 2005. The compute cost is real, but it’s a temporary condition, not a permanent constraint. Every few months, the major model providers cut prices. Processing that would have cost $9 per query in 2024 costs $0.07 now. That trend is not slowing down.
In a few years, most people won’t think about token costs any more than they think about texting costs.
What My Context Profile Looks Like
My personal context profile — the document I load into my most important agents — is 20 pages. It covers my working style, my communication preferences, my priorities, key relationships, how I like to make decisions, and dozens of other details.
I’ve seen the CEO of Lindy’s version. It’s even longer.
Neither of us has noticed meaningful performance degradation from loading it. Neither of us stresses about the token cost.
Does a longer profile cost more? Yes. But “a little more” on something that already costs very little is still very little.
The context profile is what makes an agent actually feel like it knows you. Without it, you’re prompting a stranger every time. With it, you’re working with something that has real context about who you are and what matters to you.
The ROI on that is hard to argue with.
When Cost Actually Matters
I don’t want to pretend tokens are free. There’s a real scenario where cost optimization matters a lot.
When you’re running thousands of operations at scale, a small per-operation cost compounds fast. If you’re enriching a list of 5,000 leads and each record costs $2 to process, that’s $10,000. If you can get the same quality at $0.07 per record with better architecture, you just saved almost $10,000.
That’s worth optimizing.
The rule I use: when you’re doing something thousands of times, optimize hard for cost. When you’re doing something daily or handling your own workflows, don’t bother. The marginal cost of your time thinking about it exceeds the savings.
What to Focus On Instead
The question I try to redirect people toward is this: what’s the highest-leverage thing you could automate next?
Not “how do I spend fewer tokens on this?” but “what task is eating my time every week that an agent could handle—so I can reclaim my calendar?
Those are very different questions, and they lead to very different places.
The first one keeps you in optimization mode, squeezing drops out of a bucket that’s getting cheaper every month anyway.
The second one keeps you in growth mode, finding new leverage every time.
If you’re at the stage where you’re worried about token costs on your personal agents, you’re probably not yet at the stage where it actually matters. Focus on building first. Optimize when scale demands it.
We’re still early. The text plan era of AI is temporary. Use that window to learn fast and build fast.
The costs will sort themselves out.
Want a practical framework for knowing what to automate first? The 4-Day AI Sprint walks you through it — four days of high-leverage AI skills.
