Are AI Tools Getting More Expensive — or Are We Just Using Them for Everything?
I've been wondering about something lately.
Are AI tools becoming more expensive? Or are we simply becoming more dependent on them?
Not long ago, a $20/month AI subscription felt almost unlimited. You'd ask some questions, generate some boilerplate, and end the month wondering if you'd used it enough to justify the cost. Today, I regularly hear developers — good, cost-conscious developers — saying things like:
- ▸"How did I already hit my limit?"
- ▸"Where did my credits go?"
- ▸"I used an entire month's quota in a few days."
The instinctive explanation is that the tools got more expensive — that providers reeled us in with generous limits and are now tightening the tap. I don't think that's what happened, or at least not the main thing. I think something more interesting did, and it says more about us than about the pricing pages.
Look at what we're actually asking for now
Take coding tools as the clearest example. What started as:
"Help me write this function"
...has quietly evolved into:
- ▸Design the architecture
- ▸Generate the feature
- ▸Refactor the code
- ▸Write the tests
- ▸Review the PR
- ▸Explain the bug
- ▸Debug the error
- ▸Generate the documentation
That's not the same product being used harder. That's a fundamentally different relationship with the tool. Two years ago an AI assistant touched maybe five percent of a developer's workflow — the annoying parts, the boilerplate. Today, for many of us, it participates in every phase of the development lifecycle. A single feature might involve dozens of AI interactions, each one consuming context, each one carrying the whole codebase's worth of tokens along for the ride.
The context is the multiplier people miss. "Help me write this function" was a few hundred tokens in, a few hundred out. "Review this PR against our architecture" means the diff, the surrounding files, the conventions doc, and the conversation history — thousands of tokens of context per request, re-sent with every follow-up question. Our individual asks didn't just get more frequent. Each one got an order of magnitude heavier, because we learned that AI with rich context is dramatically more useful than AI without it. Better results, bigger bills — the same lever moves both.
The per-token price of AI has actually fallen dramatically, year after year. The bill grew anyway. Which means the explanation has to be on the demand side.
We've seen this exact movie before
It reminds me of cloud computing, beat for beat.
At first, companies moved to the cloud and saved money — because they stopped buying servers they barely utilized. The finance team was thrilled. Then something unexpected happened: freed from procurement cycles and capacity planning, teams discovered they could build ten times more things. Spin up an environment for every experiment. Add a service for every idea. The friction was gone, so the volume exploded.
The bill didn't grow because cloud got expensive. Cloud got cheaper per unit the entire time. The bill grew because usage exploded — because the technology became so useful that consumption expanded faster than prices fell.
Economists have a name for this pattern — Jevons paradox: make a resource more efficient to use, and total consumption of it goes up, not down. Cheaper compute meant more compute. Cheaper AI means more AI. The $20 subscription didn't shrink. Our ambitions outgrew it.
Dependence is what utility looks like
Here's the reframe that changed my mind about the "too expensive" complaint.
Nobody complains about hitting the limits of a tool they don't need. Nobody has ever grumbled about exhausting their quota on a product they could live without — they just stop using it. The complaints are, perversely, the strongest evidence of value: developers are burning through quotas because the AI has become load-bearing in their daily workflow, woven into how they design, build, debug, and ship. It's not a toy being overused — it's infrastructure being used, in the boring, essential way electricity is used.
Maybe AI isn't becoming expensive. Maybe we're reaching the point where it has become part of our daily workflow — and we're noticing, for the first time and with some discomfort, how much we actually rely on it. The quota anxiety isn't a pricing problem. It's a dependence milestone, dressed up as one.
The question that comes next
None of this means costs don't matter — the opposite. Cloud taught us that lesson too. The usage explosion eventually produced FinOps: a whole discipline for asking "are we spending this well?" instead of just "are we spending less?" The same reckoning is arriving for AI, and the teams that handle it gracefully will treat it as an engineering problem — matching the size of the model to the size of the task, reusing context instead of re-sending it, knowing which work actually benefits from AI and which is habit.
But that's optimization, not retreat. I haven't met a single developer who hit their quota and concluded they should go back to working without AI. They conclude they need a bigger quota. That, more than any benchmark, tells you what these tools have become.
So I'll leave the question open, because I'm genuinely curious what others think:
Are AI tools getting more expensive?
Or has AI become so useful that we're consuming it faster than ever before — and the bill is just the first place we noticed?