Questions this answers
- How much should a small business expect to spend on AI?
- Why is my bill higher than the estimate I was given?
- Should I pay per seat or per token?
- Which model tier does my work actually need?
- How do I tell whether a project pays back?
AI pricing confuses people for a structural reason: providers bill in a unit nobody thinks in, businesses budget in a unit nobody bills in, and the conversion between them depends on details that only become visible after you have already committed. This page is that conversion, plus the costs that never appear on a rate card and usually turn out to be the larger number.
There are two ways to pay, and they behave nothing alike
Per-seat subscription
- A flat monthly fee per person
- Predictable, and it caps out
- Largely indifferent to how hard someone uses it
- Buys a finished product, not just a model
- Wasted on people who use it twice a month
Metered API access
- You pay for text in and text out
- Starts near zero and tracks usage exactly
- The only option for anything automated
- Lets you route cheap work to cheap models
- Unbounded — a runaway process bills happily
Most businesses that use both end up splitting: subscriptions for the handful of people who live in the tool all day, metered access for everything that runs without a person watching. That is usually cheaper than committing to either, and the crossover point is worth calculating rather than guessing — it moves a long way depending on how heavily your team actually works.
The billing unit, and why the shape of your work matters more than the size
Metered billing is per token — roughly three-quarters of a word — and input and output are priced separately. Output always costs several times more than input, at every provider. That single asymmetry drives more of your bill than the choice of model does.
It means two jobs with identical token totals can differ several-fold in cost. Summarising a long report is cheap: enormous input, small output. Generating long documents from a short brief is expensive: the ratio runs the other way. If you want one cost-control habit, it is asking for shorter outputs — which usually improves the writing anyway.
Why the real bill beats the estimate
Three effects, and almost every surprise is one of them:
Conversations re-send everything
The model has no memory, so continuing a conversation means re-sending the whole history on every turn. Turn ten pays for turns one through nine again. A long exchange does not cost the number of messages — it costs closer to the sum of a growing series.
Standing instructions bill every time
The system prompt, the pasted context, the retrieved documents — all re-sent on every single call rather than once at the start. A generous system prompt is a fixed cost multiplied by your entire call volume.
Failures bill exactly like successes
Rejected drafts, retries, and generations that came back wrong all cost full price. During development, prompt iteration burns real money on work you throw away.
Paying for capability you do not need
The spread between the cheapest and most expensive current models is roughly an order of magnitude. The spread in capability on a well-defined task is frequently close to nothing. Classifying an email, pulling a date off an invoice, routing a support ticket — none of that needs a frontier model, and running one for those jobs is the most common avoidable line on an AI bill.
The sequence that lands on the right tier without guesswork:
- Start on the balanced tier — it handles the overwhelming majority of business work
- Run the cheap tier against the same twenty real examples and compare the outputs honestly
- Escalate only the specific tasks that actually fail, not the whole workload
- Re-test when models change, because capability moves down the price curve constantly and nobody emails you about it
The costs that are not on any price list
Token cost is usually not the largest number in an AI project, and anyone presenting it as the whole cost is either inexperienced or selling something. Set against it:
- Building and integrating the thing, which in year one routinely dwarfs the running cost
- Human review time — a permanent line item, not a transitional one, because someone checks the output forever
- The management overhead of a system somebody has to own after launch
- Change: the process has to actually shift, and the hours spent getting people to work differently are real
- The quiet cost of a project nobody owns, which is how most of these die
Whether it pays back — the honest version
Most return-on-investment cases for AI multiply hours saved by an hourly rate and present the product as money. It is not money. A freed hour only becomes money if someone is no longer employed, or if that hour gets refilled with work that generates revenue. Neither happens automatically, and the first chunk of any hour freed dissolves into the day.
A defensible estimate applies two discounts the optimistic version skips: how much of the task is genuinely automatable — almost never all of it, because someone still checks and owns the result — and how much of the freed time actually turns into anything. A case that only works with both dials at maximum is not a good case. It is a case with no margin for the world being ordinary.
How much does AI cost for a small business per month?
For a small team doing ordinary drafting and summarising through the API, monthly costs typically land in the low tens of dollars — genuinely less than most software subscriptions. Per-seat subscriptions for the same team cost more but are fixed. Costs rise sharply with agent workloads and long documents, because both re-send large amounts of context on every step. Run your own numbers rather than trusting someone else's benchmark.
Is the API cheaper than a ChatGPT or Claude subscription?
It depends entirely on usage intensity. For light or occasional users the API is dramatically cheaper, often by an order of magnitude, because you pay only for what you use. For heavy daily users the subscription usually wins, since it caps out while metered billing keeps climbing. The crossover is worth calculating for your specific team size and workload.
What is the biggest hidden cost of an AI project?
Human review, because it never ends. Every other cost is either one-off (building it) or scales predictably (tokens). The requirement that a person reads and owns the output is permanent, and it is the line most business cases quietly omit — which is also why projects that assume full automation disappoint.
How do I stop AI costs running away?
Set hard spend limits at the provider before you need them. Cap conversation history rather than letting it grow unbounded. Route simple tasks to cheap models. Ask for shorter outputs. Use batch pricing for anything not time-sensitive, and prompt caching if your provider offers it — repeated context gets substantially cheaper.
Is open-source AI cheaper than paying a provider?
The licence is free; running it is not. You trade a per-token bill for hardware, hosting, and the engineering hours to keep it working well. At low and moderate volumes, commercial APIs usually win once you count salary time. The economics tip at genuinely high, steady volume — or when the reason for self-hosting was never cost in the first place, but data residency.
What to carry into a budget
- Output tokens cost several times input tokens — the shape of a workload matters more than its size.
- Conversations re-send their whole history every turn, which is the usual reason a bill beats its estimate.
- Start on the balanced tier and escalate only what actually fails. Most work does not need the frontier.
- Token cost is rarely the biggest number. Integration and permanent human review usually are.
- Hours saved are not money until someone leaves or the time gets refilled with revenue work.
Everything we have written on this
- What Is a Token, Really? The Unit That Decides AI Cost, Memory, and Mistakes Tokens are how AI reads and writes text. Once you understand what they are, pricing, memory limits, and why AI gets expensive in long conversations all start to make sense.
- How to Choose the Right AI Tool for Your Business Most "best AI tools" lists are useless for small businesses. Here's a practical framework for choosing tools that work for your specific situation.
- Free vs. Paid AI: What You Actually Get for the Money What separates free AI tiers from paid ones: data handling, reliability, limits, and the contractual difference that decides whether you can use it for client work.
- Open Source AI: What It Means and Whether It Matters for You "Open source AI" means something specific and different from open source software. Here's what it actually means, which models qualify, and whether it changes what you should use.
- AI vs. Hiring: When to Automate and When to Put a Human in the Seat Should you hire someone or set up an AI tool? The answer depends entirely on what kind of work you need done. Here's a straight guide.
- The Hidden Cost of 'I'll Do It Myself' (And When It's Time to Let AI Help) Most business owners wear too many hats and know it. Here's how to figure out which tasks are quietly draining your time, and where AI can step in.
The structured version
Same ground as a free course — objectives, sequence, and a record of what you have finished.
- What AI Actually Costs Per-seat subscriptions or per-token API billing — which is cheaper depends on two numbers you already know. Plus the real costs nobody quotes you.
- Measuring Whether It Actually Worked Without measurement, AI adoption drifts one of two ways: abandoned because nobody can show it works, or kept with nobody accountable for whether it does.
- Choosing Your First AI Project Most first AI projects fail because they were chosen to be impressive rather than to be finishable. A scoring method, a success test, and the traps to avoid.
Tools for this
- AI Cost Calculator Work out your real monthly AI bill before you commit to anything.
- Token Counter & Context Window Checker Check whether your document fits — before the model tells you it does not.
- AI Automation ROI Calculator Find out if automating a process is worth it before you spend anything.