You’re not buying a brand of AI. You’re buying a model and a meter.
We priced one month of our own AI work on six platforms. The bill ran from $42 to more than $500, and the cheapest wasn’t the one to buy. Here’s what your AI cost really depends on.
You wouldn’t buy a truck by the logo
You wouldn’t. You’d ask what’s under the hood, what it burns, and what it costs to run for a year.
Then somebody says “we should get ChatGPT,” or “we’re a Copilot shop,” and that’s the whole evaluation. A brand and a $20 price.
That’s backwards. The brand doesn’t do the work. A model does the work, and a meter decides the bill.
Most businesses have no way to price AI
They have a way to compare subscriptions. That’s not the same thing, and the difference shows up on the invoice.
The $20 or $30 a month plan is a seat. One person, typing questions into a box.
The moment AI starts doing work inside your systems on its own (sorting the inbox, drafting the follow-up, pulling the report), it stops billing by the seat. It bills by usage, on a price list most owners have never seen.
Nobody hides that list. It just isn’t on the page with the $20.
What $42 to $500 was buying
The workload was ours. Cybertitans runs AI inside the business every day: about 1,000 tasks a month. Research, drafts, inbox triage, reports.
I took that one month of work and priced it on six platforms, at each vendor’s published rates. Then I put an independent skill score next to each bill.
Four of the six make their own model.
Skill score 41
Skill score 50
Skill score 46
Skill score 51
The other two, Microsoft Copilot Studio and Perplexity, are platforms that run other companies’ models and bill in credits. Same work: $170 to $520 on one, $150 to $700 on the other, depending on how many steps each task takes.
The skill score comes from Artificial Analysis, an independent lab that runs every model through the same ten tests of real work. Higher is better. The best score anywhere right now is 58.1
Read the list again. Price and skill don’t line up.
One model costs $43 a month more than another and scores four points lower.
The cheapest one scored lowest. It’s fine for simple, bounded jobs. It’s the weakest of the four on long, multi-step work.
The top two are a tie on skill. One of them is about half the other’s bill.
That’s the tell. A price tag tells you what the meter charges. It doesn’t tell you whether the job gets done.
The four questions nobody asked
Which model is actually doing the work? The brand isn’t the model. On one vendor’s platform, picking a different model moved the same month of work from $80 to $357. And a platform that resells someone else’s model may not be selling the current one. One of the six offered an older version of a model than its maker sells direct.2
How does it bill: by seat, by usage, or by credit? Credits sound simple. On one platform every step an AI takes costs a nickel, however small the step.3 A task that takes ten steps costs three times what a three-step task does. That’s the whole gap between $170 and $520.
What does one finished task cost? Two of the “cheap” models use about twice the words to finish the same job, sometimes more. Cheap per word. Not cheap per job.
What have we already got that we haven’t tuned? The biggest saving I found wasn’t on anybody’s price list. Trimming what we send the model on every request takes the modeled bill from $165 to about $128. Same model. Same answers.
Ask a vendor which AI to buy and you’ll get theirs
That’s not a scam. It’s a menu.
Microsoft will sell you Copilot. OpenAI will sell you ChatGPT. Your software vendor will sell you the AI button inside their app.
Each of them is answering one question: which of ours? Nobody in the room is paid to answer the other one: which model fits the work, and what will the meter read in month six?
Which is why the question can’t be asked of the person selling the seat.
How I ran it
Not a recommendation. An order of operations.
- Write the work down first. What tasks, how many a month, how many steps each. If you can’t describe the work, nobody should be pricing it yet.
- Price the same work everywhere. Published rates, same tasks, same volume. Not the demo. The month.
- Put an independent score next to each price. Somebody else’s test, not the vendor’s.
- Tune what you already run before you switch. It’s the only saving with no migration attached.
- Test the best-looking alternative side by side. Your real work, scored blind.
- Build so that switching is a setting, not a project.
- Then decide.
Run in that order, here’s where we landed. We’re staying where we are for now. We’re tuning first. And the model that looks like the best value on paper gets a side-by-side on our lowest-risk work before anything moves.
Where this came from, and what my stake is. This one is our own house, not a client’s. The workload is Cybertitans’ and so is the bill.
We run on Claude today, so I’m the incumbent’s customer. The research was also run with Claude, which is one of the six being compared. Every price was read off the vendor’s own page and every score off an independent lab, so you can check them without taking my word or its word.
I’m a Microsoft partner and I resell Microsoft licensing. The Copilot Studio numbers aren’t kind, and I’ve left them in.
And managing AI for clients is something Cybertitans sells. “Somebody should own this” is a conclusion I make money from. Weigh the argument on its merits rather than on my neutrality, because I don’t have any.
What hasn’t happened yet. Every figure here is modeled from one month of our workload at published rates on September 30, 2026. Nothing has been saved. The tuning isn’t done and the side-by-side hasn’t run. A piece about price tags overstating what you get isn’t going to overstate what happened.
The real question isn’t the price. It’s who’s watching the meter.
AI pricing moves faster than anything else on your invoice.
Three of the four models on that list shipped in the last ten days. One vendor’s price is set to double on January 1.4 Another retired eight models in a single day this spring and moved its customers onto one with a different price.5
So whoever set your AI up in March isn’t looking at it in September. The plan nobody’s watching is the expensive one.
And the meter isn’t the only thing nobody’s watching. Many consumer AI plans can use what your people type in to train the model. That’s not a pricing problem. It’s a policy problem, and it’s why we published a free AI Acceptable Use Policy alongside this.
Seven questions to ask before AI goes in the budget
If a vendor can’t answer these in a first meeting, that’s your answer.
- Which model will do the work, and is it the maker’s current one?
- Is this billed by seat, by usage, or by credit?
- What happens to the bill when the work doubles?
- What does one finished task cost, not one seat?
- What happens to what we put in? Is it used for training, how long is it kept, can we delete it?
- If a better model ships next month, what does it take to switch?
- Who on our side checks the bill and the output every month?
None of it is technical. All of it is answerable in a first meeting. Most vendors have never been asked.
Does this sound like something happening in your organization?
If AI is about to go into next year’s budget as a line that says “$30 a seat,” that’s the conversation to have before the purchase order. Not after.
This is the work: sitting on your side of the table while vendors sell to you, and staying long enough to be accountable for whether the call was right.
Book 20 Minutes- Artificial Analysis, Intelligence Index, read September 30, 2026. Leaderboard
- Microsoft Learn, models available in Copilot Studio. Select an agent model
- Microsoft Learn, Copilot Studio billing rates. Billing rates and management
- Google, Gemini API pricing: introductory prices through December 31, 2026, doubling January 1, 2027. Pricing
- xAI, retirement of eight model IDs on May 15, 2026, with requests redirected to a differently priced model. Migration guide
- Published rates used for the cost model: Anthropic · OpenAI · xAI · Google · Perplexity credits