Plain language

What AI actually costs a small business

August 2, 2026 · 8 min read

The most common question I get isn't what AI can do. It's what it's going to cost. Nobody answers that straight, because the honest answer has two halves and the people selling it only like talking about one. Here are both halves, with real numbers — and the numbers on this page get re-checked against the vendors' own pricing pages every month.

Rates verified August 2, 2026 against the published pricing pages listed at the bottom. If a rate on this page is stale, it's because a vendor changed it in the last few weeks — everything here is checked monthly and the date above moves when it is.

Half one: seats

What you pay for people to use the chat tools directly.

This is the half everyone understands, because it works like every other piece of software you buy. A per-person monthly fee, billed like your email or your accounting package. Across the major business tiers it lands in the same neighborhood — roughly the cost of a couple of lunches per person per month. Check the specific vendor's page for the current figure, because unlike the usage rates below, seat prices are quoted differently for annual versus monthly and for different tiers.

Two things worth knowing before you buy seats:

The business tier is not just about features. The consumer and business tiers of the same product often handle your data differently. That's usually the real reason to upgrade, not the extra capacity. I wrote about where that line sits in shadow IT versus sanctioned tools.
Seats are cheap; adoption is expensive. Licenses handed out without a specific job to do with them get opened twice and abandoned. The cost that actually hurts is the month you paid for twenty seats and three people used them.

Half two: usage

What you pay when AI is wired into a workflow rather than typed into a chat box.

This is the half nobody explains, and it's where the surprises live. When a script calls an AI model — to sort an inbox, read an invoice, categorize an order note — you're billed by the amount of text going in and coming out, measured in tokens. A token is roughly three-quarters of a word. A page of text is about 500 tokens. Prices are quoted per million tokens, which is why they look reassuringly small and why people underestimate them.

What the major models cost right now

Per million tokens, in US dollars, for the standard API rate:

Model	Input	Output
Claude Haiku 4.5	$1.00	$5.00
Claude Sonnet 5	$3.00	$15.00
Claude Opus 5	$5.00	$25.00
Claude Fable 5	$10.00	$50.00
GPT-5.6 Luna	$0.20	$1.20
GPT-5.6 Sol	$5.00	$30.00
GPT-5.5 Pro	$30.00	$180.00
Gemini 3.5 Flash-Lite	$0.30	$2.50
Gemini 3.6 Flash	$1.50	$7.50

Look at the spread before you look at any individual row. The cheapest option on this table costs about one-fiftieth of the most expensive one for the same amount of text. That gap — not the absolute numbers — is the thing to take away. The difference between a thoughtful setup and a careless one isn't ten percent on your bill. It's most of your bill.

Also note that output costs three to six times what input costs, everywhere, without exception. Reading is cheap; writing is expensive. A workflow that reads a lot and writes a little is the cheap kind.

What that means in actual work

Numbers per million tokens are meaningless until you can picture the job. So: summarizing a thousand customer emails. Call it 400 words per email going in and a 60-word summary coming out. That's roughly 500,000 tokens of input and 80,000 of output.

On a small, cheap model: about a dollar.
On a mid-tier model: about two to three dollars.
On a flagship: about four to five dollars.

A thousand emails. For less than the cost of lunch, on any of them. That's the part that surprises people in the good direction, and it's why "the AI bill" is almost never the reason a project doesn't work.

Now run the same job every day for a year on the flagship when the cheap model would have done it, and you've spent roughly fifteen hundred dollars to buy judgment you didn't need. That's the part that surprises people in the bad direction. Neither number is big. The ratio between them is.

Two discounts worth knowing about

Both are standard across the major vendors, and both are easy to leave on the table:

Batch processing: about half price. If a job doesn't need an answer this second — an overnight report, a weekly cleanup pass — you can submit it as a batch and pay roughly 50% less. Most small-business automation is exactly this shape and gets run at full price out of habit.
Caching repeated input: up to 90% off. If every call sends the same long block of instructions or reference material, that repeated part can be cached and billed at a fraction. This is a build detail rather than a setting you toggle, but it's the difference between a workflow that costs what you expected and one that costs five times that.

The choices that actually move the number

In order of how much they matter:

Whether you use a model at all. If a task follows fixed rules — this column maps to that field, this status routes there — a plain script does it for free, forever, identically every time. Calling a model on every row is a tax on work that never needed judgment. This is the single biggest lever and it's the one most people skip; I made the whole case in no token tax.
Which model you point it at. Shallow judgment — classify this, extract that, reformat this — belongs on a small model. Deep judgment belongs on a flagship. Getting this wrong is the order-of-magnitude mistake, and the model-selection map is how to get it right.
How much text you send. Sending an entire document when three paragraphs would do multiplies your input cost by whatever the ratio is. Long chat histories do this quietly, because the whole conversation is re-sent every turn.
Whether you batch and cache. See above. Worth real money on anything running at volume.

The pattern underneath all four is the same one that shows up in every workflow that actually ships: a scripted spine doing the deterministic work for free, with a narrow AI call at the one joint that genuinely needs judgment.

The cost nobody puts on the invoice

Everything above is the easy half to calculate, and it's rarely the number that decides whether the project was worth it. The costs that actually bite:

The time to describe the process precisely. This is the real work, and it's yours — nobody can do it for you, because you're the one who knows the exceptions. It's also most of the value, which is why I keep saying so.
Checking the output until you trust it. Every AI workflow needs a verification period. Budget for it, because the alternative is trusting a confident wrong answer.
Somebody owning it. An automation nobody maintains quietly stops matching reality about six months in.

Against those, the token bill is usually a rounding error. I've never had a small business project where the model cost was the deciding factor. Not once.

If you're deciding whether to try AI on something, the cost question is almost never the one that matters. Run the numbers, confirm they're small, and then go spend your attention on whether the task is a good fit at all.

The takeaway

Two halves: seats, which cost about what any other software seat costs, and usage, which is priced per million tokens and spans a fifty-fold range depending on which model you pick. Both are smaller than people expect. The expensive mistakes are using a model where a free script would do, and using a flagship where a small model would do — not the bill itself.

Sources

Every rate on this page comes from the vendor's own published pricing, checked on the verification date at the top:

Model names and prices in this market change every few weeks. If you're reading this well after the verification date and the decision matters, click through and check.