AI is getting cheaper, and not by a little.
When the first genuinely capable models arrived a few years ago, running them cost enough that you thought twice about each use. The price of running the same quality of work has fallen by something like ninety per cent in two years, and by far more than that against those earliest models. The cheapest capable models now read a paragraph and write a short reply back for a fraction of a penny. Epoch AI, which tracks this carefully, found that the price to reach a given level of capability has fallen somewhere between roughly nine and nine hundred times a year depending on the task. That is a real and fast collapse, and you have probably been told what it means: AI is basically free now, so automate everything.
That is the wrong lesson. Not because the prices are not falling, but because the running cost was rarely the thing standing in the way.
AI processing costs have fallen
The price that collapsed is the price of the raw run: the few pennies it costs to have a model read something and write something back. That genuinely matters in one place, which we will get to. But for most of the automations a small business is weighing up, the per-run price was never the line that killed the idea. It was a rounding error next to the time someone spent setting the thing up, checking what it produced, and keeping it alive afterwards.
So the headline number has fallen dramatically, and most of the real decision has not moved at all.
Small repeated tasks now cost less to automate
There is a real category that cheaper models unlock, and it is worth naming precisely so you can spot it.
It is the high-volume, low-stakes job where each single run is nearly worthless but you do it thousands of times. When each run cost real money, the sums never worked, because the cost multiplied by the volume was more than the job was worth. Now it does work.
We built an AI feature for an art gallery shop that looks at each incoming artwork image and drafts a title, a description, alt text, and suggested tags. Done across a whole catalogue, image by image, that is exactly the kind of repetitive, low-drama task that cheap models make affordable and that a couple of years ago would have been borderline. Tag every product image. Draft alt text for every page. Give every incoming transaction a first-pass category. Normalise every messy supplier line into the same format. The job itself did not change. The price of doing it ten thousand times did, and that is the genuine shift.
Review and ownership still cost the same
Now the catch, because it is the part the headline skips.
Picture a small business getting sixty enquiries a week through its website form. Having a model sort each one and draft a first reply now costs pennies a week to run. So the question "can we afford to run this" is answered: yes, easily. But two other costs are sitting exactly where they were.
The first is review. Someone still reads those drafts before they go to a customer. If the model drafts sixty replies and a person checks sixty replies, you have moved the work, not removed it, and the running cost was never the expensive part of that work. The cheaper the run gets, the more obvious it becomes that the review is the real budget.
The second is ownership. When the form changes, when a better model comes along and you swap to it, when the thing quietly starts doing something slightly odd, someone owns that. Cheap to run is not the same as cheap to own, and the part that costs you sits in the years after launch, not the afternoon you set it up. The price of the run fell. The price of keeping a working tool alive did not.
Volume can erase small unit savings
The "fallen ninety per cent" figure is true for the easy end of the scale. Prices dropped furthest and fastest for short, simple tasks. They fell much less for the hard end: a model that reasons for a long time, reads your whole knowledge base on every question, or works like an agent grinding away at a job for an hour.
So "AI is basically free" is true for a one-line classification and untrue for the heavy work where the bill actually shows up. The awkward part is that the tasks most worth automating in a small business often sit in the middle, useful enough to need a capable model, repeated enough to add up. The falling price helps there, but it does not make the decision for you.
Cheaper AI changes the edges, not the decision
For most decisions, less than the headline suggests. The question moved, but only a step. In the handful of cases where running cost ever mattered, it used to be "can we afford to run this". Now it is almost always "can we afford to review it and keep it going".
Before you automate something because it is suddenly cheap to run, check the three things that actually decide it, and notice that only one of them got cheaper:
- Does the task repeat often enough to be worth wiring up at all? Unchanged.
- Can a person review the output without the review costing more than the task saved? Unchanged.
- Is someone going to own it when it drifts, or the form changes, or you switch models? Unchanged.
Our own scheduled agents, the ones that sort incoming work, draft the Monday-morning project summary, and group live software errors for us, earn their place on whether the task repeats and whether the output can be checked, never on what a single run costs, because the per-run cost was almost never the thing in the way.
The useful version of "AI got cheaper" is not "automate everything". It is narrower and more practical: a few more high-volume, low-stakes jobs are now worth doing, and everything else is decided by the same questions as before. If the only thing that has changed about a task is the price tag on the run, that is usually not enough on its own to act on.
If you are looking at a job and trying to work out whether it is one of the few the falling price genuinely unlocked, or one where the real cost was always somewhere else, that is a good thing to settle before you build anything. The next step is our Free 20-minute call. We will tell you which one it is.
