What It Costs, and Keeping It Near Zero

Four shapes of price

Free tiers. ChatGPT, Claude, Gemini, Copilot, DeepSeek, Mistral and others all have one. Real work fits inside them: capped daily messages, sometimes older models. Many small businesses never need to leave.

A flat subscription. Roughly USD 20 a month per person for the consumer paid tiers, with regional pricing in some markets: around ₹1,700 to 2,000 in India, for example. Prices and limits change often enough that you should check rather than trust this paragraph.

Pay per use (the API). You pay by the token, roughly a word-and-a-bit. A page-in, half-a-page-out exchange costs about one US cent on a mid-priced model and closer to a tenth of a cent on a cheap one. Five hundred customer replies a month lands somewhere between USD 0.50 and USD 6. The catch: an API key is not an app. Something has to call it.

Bundled into software you already buy. Your accounting package, your CRM, your helpdesk. Often the cheapest real option, because the data is already there.

Compare against the thing it replaces

This is the whole discipline of this lesson.

USD 20 a month is expensive against a free tier that was doing the job. It is cheap against six hours of your own time. It is absurd next to a USD 4 scheduling tool that solved the same problem completely.

So the question is never “is this cheap?” It is “cheaper than what, exactly?” Name the alternative before you name the price.

Where the money actually leaks

Four patterns, in order of how often they catch people:

Stacking. A writing tool, a social tool, a meeting tool, a chatbot. Four at USD 20 is USD 960 a year, and one general assistant plus a facts sheet did most of it. Audit your list twice a year.

Per-seat creep. Team plans multiply. Five people at USD 25 to 30 is USD 1,500 to 1,800 a year, and often two of those five open it monthly.

Annual plans bought on day two. The discount is real. So is the fact that you bought twelve months before you knew whether it worked. Monthly until you have run the two weeks in lesson eight.

Credits that expire. “1,000 AI credits included” in software you already pay for. Find out what a credit is, whether it rolls over, and what the overage rate is, before you build a habit on it.

Keeping it near zero, in order

1. Start free. Genuinely. A month on a free tier tells you whether the habit sticks.
2. Check what you already pay for. There is a decent chance two tools on your bank statement have AI features you have never opened.
3. One paid tool at a time. Add the second only when the first has clearly earned its place.
4. Cap the API. If you go the pay-per-use route, set a hard monthly spend limit on the key. The major providers all support this. A runaway script is otherwise a memorable bill.
5. Consider local models. Llama, Mistral, Qwen, Gemma and others run on your own machine through tools like Ollama or LM Studio. If you already have a reasonably modern computer with enough memory, marginal cost is zero and nothing leaves the building, which also solves half of lesson seven. Quality is a step below the best hosted models, setup is an afternoon, and it is a good fit for repetitive sorting and drafting.

A realistic monthly figure

For most one-to-five person businesses doing the two tasks from lesson one, the honest number is USD 0 to 25 a month, all in.

If you are budgeting USD 200 a month for AI in a business of three people, something has gone wrong upstream: usually stacking, usually because each tool was bought without asking what it replaced.

And be suspicious of any tool priced at a percentage of your revenue or per customer contact. That pricing scales with your success rather than with its cost, and small businesses are where that model does the most damage.