What an AI Stack Actually Costs a One-Person Business

Nobody subscribes to five AI tools. They subscribe to one tool five times, in slightly different clothes, and find out at the end of the quarter.

This post is the arithmetic I wish someone had shown me before my card did. The numbers below are current as of August 2026 and checked against multiple sources — but prices in this industry change like weather, so treat the method as the durable part and re-check any number before you act on it.

The baseline stack everyone drifts into

A solo operator who works in text and code tends to accumulate some version of this:

Tool	Tier	Price/mo
ChatGPT	Plus	$20
Claude	Pro	$20 (or ~$17/mo billed annually)
GitHub Copilot	Pro	$10
Cursor	Pro	$20

That’s $70 a month — $840 a year — and it’s the restrained version. No image tool, no voice tool, no transcription, no “it was $8 and I forgot it exists.” Most stacks I’ve seen are north of $100 once those are counted.

$840 isn’t the problem. $840 for capability you’re paying for twice is.

Where the overlap hides

The uncomfortable question for every line in that table: what does this do that another line doesn’t?

Two chat assistants. ChatGPT Plus and Claude Pro overlap on most everyday work — drafting, summarizing, thinking out loud. Keeping both is defensible only if you can name the task where each one wins for you, from your own use, not from a benchmark chart. If you can’t name it, you have a $240/year tiebreaker subscription.
Two coding tools. Copilot autocompletes inside your editor; Cursor is an editor with the model wired through it. They solve the same hour of your day. Running both usually means one is a habit, not a tool.
A chat subscription doing a specialist’s job. Before adding a $20 “AI for X” tool, spend ten minutes checking whether the chat assistant you already pay for does X at 90%. The specialist earns its slot only by wrapping a workflow or living where the output lands — the test from How to evaluate an AI tool before you pay.

The stack rule: one anchor, one specialist, one experiment

The shape that survives contact with a bank statement:

One anchor — a general assistant you use daily and know deeply. Depth beats breadth: prompt habits, saved context, and knowing the tool’s failure modes compound with one tool and reset to zero with each switch.
One specialist — the tool for the job that actually earns your money (for most solo builders, the coding tool). It must beat the anchor at that job by enough to notice, on your own work.
One experiment, time-boxed — a rotating slot for whatever you’re evaluating, with a calendar reminder to decide at the end of the month. The slot is permanent; its occupant never is.

Three slots, roughly $50/month. Everything else has to evict something to get in — which forces the comparison that ad-hoc subscribing never does.

The audit takes fifteen minutes

Once a quarter, list every AI charge from your actual statement — not from memory, memory is exactly what fails here. For each line, write one sentence: “Last week I used this to ____.” Past tense, specific, last week.

A line you can’t complete is a cancellation, not a debate. If the sentence is about what you plan to use it for, that’s a plan, and plans are free on the free tier. (What cancelling actually costs you — and how to test it safely — is the subject of The cancel test.)

The one-page version

A restrained text-and-code stack runs ~$70/mo in August 2026; typical real stacks exceed $100.
For every subscription, name what it does that no other line does.
One anchor, one specialist, one time-boxed experiment. New tools evict.
Quarterly audit from the statement, past-tense sentence per line.
Re-check every price before acting — this industry reprices constantly.

None of this is about spending less on AI. Good tools are worth real money — paying $50/month for tools you actually drive is a better deal than $100 for a drawer of overlapping trials. The goal is a stack where every line would survive being asked what it did last week.