The Real Cost of AI: The Three Hidden Costs Your Budget Forgets
By Anis Hammouche·July 6, 2026
The real cost of an AI use has three lines: what you pay as volume goes up, the time you spend fixing its mistakes, and what depending on one provider costs you. The API bill shows only the first. In the worked example below it comes to a few dollars, while the rework is counted in hours.
Take Inès, who runs an online tea shop on her own (an invented case, like every figure in her examples in this article). She has AI draft her replies to customers. The day she looks at her API bill, she thinks it is almost free. She does not count the evenings she spends rereading. That is the trap: the only cost that shows up on an invoice is often the smallest.
Usage cost: the bill that climbs with volume
This is the visible cost. Most AI services charge by the amount of text processed, counted in tokens (a token is a small piece of text, a word or part of a word). On November 1, 2026, Anthropic's pricing page lists Claude Sonnet 5.5 at 2 dollars per million input tokens and 10 dollars per million output tokens. Generated text therefore costs five times more than the text you send. Prices move: reread the page before you rely on it.
Here is the calculation for Inès's shop, using assumptions: 1,000 replies a month, each with 2,000 tokens sent (the customer's message and your instructions) and 500 tokens generated.
| Line | Calculation | Amount |
|---|---|---|
| Input | 2 million tokens × $2 | $4 |
| Output | 0.5 million tokens × $10 | $5 |
| Monthly total | $9 |
Nine dollars a month. Multiply the volume by ten and the bill follows: 90 dollars. This cost is not a problem as long as you estimated it beforehand. The same page lists two levers: rereading text that is already cached costs 5% of the normal input price on this model (writing it to the cache the first time costs a little more), and batch processing, for anything that is not urgent, costs 50% less. A scoped use lets you take them. An assistant that everyone opens whenever they like leaves you nothing to estimate.
Rework cost: the time you spend when AI gets it wrong
This is the forgotten cost, because it is on no invoice. When AI answers wrongly, someone catches it, fixes it, sometimes redoes the work. That someone is you.
Back to Inès with another assumption: one reply in five needs a 4-minute correction. Across 1,000 replies, that is 200 corrections, 800 minutes, a little over 13 hours. At 40 dollars an hour (use your own rate, mine is not yours), that is about 530 dollars of time a month, against 9 dollars of API. Both figures are invented, but the ratio is the point: human time weighs far more than the token.
An error rate of one in five is not a fact about AI in general, it is an example. You measure your own: take 20 of your outputs, count the ones that pass without a fix, time the others. The same rework shows up for people who sell their time: a small agency that has AI write its drafts and then rereads them pays this cost without invoicing it. For more on this, read The AI Productivity Paradox.
Dependency cost: the day the provider changes the rules
The third cost does not trigger all the time, and it lands at a bad moment. When you build a use on an outside service, you accept that its rules can change. A price can move. A model can be retired. On Anthropic's deprecations page, the provider says it regularly retires older models and gives at least 60 days' notice before retiring a publicly released one. Other providers have their own rules: read theirs.
Sixty days is short if your tool serves a client. Take a freelance developer who delivers a tool built on one specific model (an invented case). When that model is retired, he has to switch to another, retest, and fix whatever changed. If nothing was planned in the quote, nobody pays for that time. The defence takes three moves: note which model each use depends on, put a maintenance line in your quote, and keep a fallback. I cover the fallback in Your AI Tool Can Be Cut Off Overnight.
The three costs side by side
| Cost | What it is | In the shop example (fictional) | First move |
|---|---|---|---|
| Usage | The bill that climbs with volume | $9 a month, $90 at ten times the volume | Estimate volume before launch |
| Rework | Your time spent fixing | About 13 hours, or 530 dollars a month | Measure on 20 real outputs |
| Dependency | A price or a model that changes | Left out of the calculation, but 60 days' notice | Note what each use depends on |
All three exist whether you look at them or not. Looking at them lets you choose which use to keep. In Inès's example, rework decides, not the token. If she corrected only one reply in ten, her gain would look very different. You can measure what a tool earns you with the same before-and-after logic.
Cost a use in ten minutes
Copy this sheet and fill it in for a use you have or are considering:
AI use: [what the AI does for me]
Monthly volume: [number of requests]
Usage cost: [price per million tokens, on the day I read it] × [tokens per request] × [volume]
Share of outputs to fix: [out of 20 real outputs, how many]
Rework time: [minutes per fix] × [number of fixes] × [my hourly rate]
Dependency: [model or service used, date I last checked the deprecation page]
Heaviest cost: [usage / rework / dependency]
Verdict: [I keep it / I narrow it / I stop]
Frequently asked questions
Is the API bill enough to budget an AI use? No. It covers only usage cost. It ignores the time you spend fixing and the dependency risk. In Inès's example it comes to 9 dollars, against about 530 dollars of rework time.
How do I measure rework cost before launching? Take 20 real outputs from the tool. Count the ones that pass without a fix and time the others. Multiply the rework time by your hourly rate. It is an order of magnitude, not a certainty, but it is yours.
Should I avoid outside providers to remove dependency cost? No, that would be out of proportion for a small business. Write down what each use depends on, read the provider's deprecation page from time to time, and plan maintenance into your quotes if you deliver a tool to a client.
Does a use that everyone opens cost more than a scoped one? That is my opinion, not a measured fact: yes, because you cannot estimate the volume, the rework or the dependencies of a use nobody tracks. A scoped use lets you cost all three lines before you pay.
Sources
- Anthropic, Pricing (Claude Sonnet 5.5 prices, caching and batch processing, read on November 1, 2026)
- Anthropic, Model deprecations (at least 60 days' notice before a publicly released model is retired, read on November 1, 2026)
The examples of Inès, the freelance developer and the agency are fictional, along with every volume, time and hourly rate attached to them.
The next step is Replacing a Job With AI: The Hidden Cost That Bites Twice, which extends this calculation to hiring. I share this kind of calculation in my newsletter, you can sign up right below.
