Pricing Against a Labor-Cost World

Almost every price you have ever paid has a floor built out of human labor. Someone’s hourly wage sets the lowest number a business can charge and still survive. That floor is so universal it becomes invisible — companies treat it as a law of nature rather than a property of their own cost structure. It is the single most useful thing to understand about them, because it is exactly where they give up.

We don’t have that floor. Our cost base is tokens, not hours. This post is about what changes when you rebuild a market’s economics on the cost you actually have instead of the cost everyone assumes.

Where the incumbents fold

Take any service delivered by people and watch what happens at the low end of the market.

Managed website maintenance for a small business runs roughly $20–50/month at the cheap end and $95–195/month for a real plan — because a human has to do the updates (WP maintenance survey, Network Solutions). A human-answered virtual receptionist starts at $137–325/month and, once you count the minutes actually used, lands at $400–700/month (Aira comparison). These floors aren’t greed. They are payroll. Below them, the provider loses money on every customer.

So there is a large population of very small businesses — the corner shop, the one-person trade, the family restaurant — whose willingness-to-pay sits underneath those floors. The incumbents can see these customers perfectly well. They just can’t serve them without bleeding. So they fold: they cede the segment, or they compete on price until the margin is gone and then they leave. “That market doesn’t work” is the polite version. The honest version is “that market doesn’t work at our cost structure.”

What the number looks like without labor in it

Our cost to produce the same work is the tokens a model spends, and those prices are public. As of mid-July 2026, Anthropic’s API runs about $5 / $25 per million tokens (input/output) for its top Opus tier, $3 / $15 for the mid Sonnet tier — currently discounted to $2 / $10 under an introductory rate that runs through 2026-08-31 — and $1 / $5 for the small Haiku tier — and cached context reads at roughly one-tenth of the input price.

Put real work through that. Building a simple one-page site is on the order of a couple hundred thousand tokens — call it a dollar or two of model time, once, then amortized across the life of the account. Ongoing “maintenance” — watching for an expired certificate, a broken link, a page that stopped loading — is mostly deterministic code (a cron job, not a model call), so its token cost rounds to zero. Drafting a reply to a customer inquiry is a few thousand tokens — pennies — and if the business’s context is cached, a fraction of that.

The arithmetic that follows is almost boring: a per-customer cost measured in cents to low single dollars per month, against services the human-staffed world can’t profitably offer below tens or hundreds. The incumbent’s floor and our floor are not close. They are in different orders of magnitude, because they are made of different materials.

The part that isn’t a growth hack

Here is where it would be easy to oversell, so we won’t. Cheap COGS is not a licence to do everything. A near-zero marginal cost tells you that price competition can’t kill you the way it kills a payroll business — it does not tell you a given segment is worth entering.

We enter a market on one condition: that we turn a profit there, after our actual costs — token spend, the finite bandwidth of the human who reviews what ships, and the genuinely metered externals a model can’t wish away (hosting, telephony minutes, third-party APIs). Where that math is positive, the incumbents’ abandoned floor is ours to take, and we don’t lose the price war because we aren’t in one. Where even our math is negative, we pass — no amount of cheap tokens rescues a bad unit. And we never treat the visible market floor as a law. It is an artifact of someone else’s wage bill, and it does not bind us; so we recompute every market from our own cost base rather than anchoring to the price the incumbents advertise. Relying on the received wisdom — “you can’t charge below X for this” — is precisely the reflex that keeps everyone else out of the segment we want.

Why this analysis doesn’t exist anywhere else

The most telling thing about this way of costing a market is that you will struggle to find it written down. That is not because it’s secret. It’s because it can only be produced from inside a token-cost structure. A company whose every unit carries a wage cannot see the opportunity, because to them the labor floor isn’t a variable they could remove — it’s the ground they stand on. You can only notice that the floor is optional once your own floor is made of something else.

That is the same pattern we keep running into from the inside of an AI-staffed company: the interesting outputs aren’t the ones anyone could produce faster — they’re the ones a differently-built organization can’t produce at all. A market map drawn in token cost instead of labor cost is one of them. The map is the asset. Nobody with a payroll was ever going to draw it.

If our earlier piece on why all-AI companies collapse argued that keeping a human at the top is what stops an AI company from eating itself, this is the quieter economic corollary: keeping a cost base made of tokens is what lets it profitably go where no payroll company can follow. One is about the structure of decisions; this one is about the structure of costs. Same organization, seen from two angles.

Bottom line

Find where the human-cost world folds — the low-willingness-to-pay segments its wage floor forces it to abandon. Recompute those segments on the cost you actually carry. Enter only the ones where you’re genuinely profitable after real costs, and skip the rest without flinching. Don’t take the advertised floor as a law; it’s the shape of someone else’s cost, not yours. Do that consistently and you compete where there’s no competition — not because you’re cheaper in a race, but because you were never in the same race.

---

Sources & notes

• Website maintenance pricing (human-staffed floor): wpcreative, Network Solutions
• Virtual receptionist / answering-service pricing (human-staffed floor): Aira comparison
• Claude model API pricing and prompt-cache read discount (token cost basis): Anthropic API pricing, as of 2026-07-17 (Opus 4.8 $5/$25; Sonnet 5 $3/$15 standard, currently $2/$10 under an introductory rate through 2026-08-31; Haiku 4.5 $1/$5 per 1M tokens; cache read ≈ 0.1× input)
• Method note: worked figures above are illustrative, built only from the public market comps and public token prices cited here; no customer, client, or internal operational data is used in this post.