Speed Made Good

The grass on the practice field was wet at dawn, and I was walking it with a phone held out in front of me, counting paces under my breath.

That was 2007. I was twenty-seven, building the prototype of a golf app on the GPS inside a BlackBerry Curve 8330, a business phone with a small keyboard under the screen. Before I trusted it on a real course, I tested it on ground I could reach. I plotted points around the field, then paced them off to see whether the phone's distance agreed with my feet.

That was checking an instrument against ground I already knew. It was useful, Emma. But it only told me how well my own instrument read. It didn't tell me whether anyone would ever play a round with the thing.

Navigators keep two speeds apart. Speed through the water is what the boat's own log measures: how fast she moves through the water around her. Speed made good is the ground she actually covers toward where she's going, after the current has had its say. The two can differ by a lot, and the log can only ever give you the first.

The manual at the head of this letter says the reckoned position is "only an approximate position" because it "does not allow for the effect of leeway, current". That's the problem with most of the numbers you'll be shown about what has got cheap. They're logs. They tell you how fast the boat moves through the water, and nothing about the current.

Here is what the record has on the speed through the water. Smith.ai, a company that answers phones for other businesses, lists people taking 300 calls a month for $2,100, and its AI taking the same calls for $500. That's its own price page, read on one day in September 2026.

App-building tools list at $16 to $25 a month on their makers' own pages, the cheapest when billed for a year.

Underneath those tools sit the models, and their price has fallen fastest of all. By the count of Epoch, a research group, the list price to reach a given score on a model benchmark "fell between 9x and 900x per year, median 50x", on data that ends in early 2025.

The fall isn't everywhere. Two captures of Google's translation page, four years apart, show the same price: $20 per million characters in 2022, and $20.00 in 2026.

These are honest logs. They tell you what the seller charges. They say nothing about what the job costs the person who needed it done.

More boats are leaving harbour too. New subscription apps went from about 2,000 a month to more than 14,700 between January 2022 and January 2026, on Appfigures' estimate as the payments company RevenueCat relays it. On Steam, the big store for PC games, an analyst counted launches flagged as made with AI going from about 13 a month, before Steam required studios to say so, to about 530.

Those are departures. They don't say how many boats arrived anywhere.

The record doesn't price the finished job. The research found no total cost of the task in any of the seven needs it checked, and no price on an owner's time anywhere.

So here's my reckoning, an estimate with its assumptions shown. Take the phones first. On the price page, the AI saves $1,600 a month. What it costs on top is the owner's own time: setting it up, reading what it wrote down, fixing what it got wrong. The research found no measure of that time, so I priced an owner's hour between two government median wages, $24.51 for all US workers and $50.85 for general and operations managers. At those rates, the AI stays cheaper until the owner spends somewhere between about 31 and 65 hours a month on it.

Reckoned all in, with a few hours of the owner's upkeep a month and a share of calls handed to a live person at $3 each, the AI route likely costs between about $525 and $1,378 a month, about $750 in the middle, against $2,100.

So here the cheap route really is cheaper, by less than the sticker says. And the break-even is a number you can hold against your own week. A machine answering a phone isn't the same product as a person doing it, though, and nothing in the record holds the quality steady between them.

Now the app. By Clutch's figures, drawn from client reviews, most agency-built mobile apps cost between $10,000 and $49,999, and the average is $90,780.

A builder tool lists at $25 a month. Here's the reckoning for a small web app built yourself, scoped as one small business's app with a login, a few tables, about five screens and a booking or payment form.

Its first year likely costs between about $682 and $20,295, around $3,716 in the middle, and that middle figure includes a $699 security scan.

Twelve months of a $25 tool is $300, under a tenth of it. Most of the rest is the builder's own hours, priced between the all-workers median and a software developer's. Leaving security reviews out of both sides, building it yourself stops being cheaper than $10,000 somewhere between about 148 and 395 hours of your own work.

An agency's build and a self-built app aren't the same product; reviews, security and upkeep differ. But the shape holds: the tool's price is speed through the water, and your hours are the current.