What an AI Tutor Really Costs, and the Case for It

Last week I sat down to price a pilot. Thirty middle schoolers, a month of AI tutoring, including live voice, the kind where a child talks to the tutor out loud and it talks back. I wanted a real number before I promised anything to thirty families, so I built a spreadsheet from the models our app actually calls and the prices the companies actually charge.

Then, out of curiosity, I put a private tutor next to it. The comparison surprised me less than the details did. The headline is that AI tutoring is cheap. The details are where it gets interesting: one setting most people have never heard of can make it ten times more expensive, and cheap AI used the wrong way can make a child worse at math. So this is the whole picture, numbers first, then the case for AI tutoring, then where a person still wins.

What a human tutor costs

In the US, a private math tutor for a middle schooler usually runs somewhere between $30 and $60 an hour, with online tutors at the lower end and big cities at the higher end (Mathnasium puts the full range at $25 to $125). Take $50 as a middle figure and one hour a week, which is what most families I know actually do.

That comes to about $217 a month, or $2,600 a year, for roughly four hours of tutoring a month.

Those four hours are worth a lot. They are also a small slice of a month of homework, and they arrive on Tuesday at five, not on Sunday night at nine when the fractions stop making sense.

What an AI tutor costs, line by line

An AI tutor is billed by the word, more or less. The companies charge per "token," a chunk of text roughly three quarters of a word long, at a few dollars per million. Here is what one middle schooler costs in a typical month on our app, using Anthropic's Claude for the thinking and OpenAI for the voice:

• A question to the tutor and its reply: about four tenths of a cent. A hundred of them a month is about 43 cents.
• Reading a photo of a homework worksheet: about 3 cents. Ten photos, 33 cents.
• Helping through each step of that homework: a third of a cent a step.
• A teach-back, where the child explains an idea and the tutor asks follow-up questions and grades it: about 4 cents.
• The tutor reading its questions aloud: a cent and a half a minute.
• Planning the week, quizzes, the lessons themselves: nothing. They do not use AI at all.

Add it up and a typical month is about $2 a child. A light month is under a dollar. A child who uses it every evening and photographs every worksheet comes to about $4.70.

Then there is live voice, which is where the money is. A real spoken conversation with an AI tutor costs between 2 and 8 cents a minute, depending on the model. Three hundred minutes a month, which is five hours, about what a one-hour-a-week human tutor gives, comes to:

• About $23 a child with the flagship voice model.
• About $9 with the smaller "mini" model.
• About $6 if you skip live voice and let the child talk through the browser's free dictation while the tutor answers in text and reads it aloud. Slower, a little stiffer, and the child cannot interrupt, but it works.

Add hosting and the rest, and thirty children with five hours of live voice each come to somewhere between $270 and $780 a month, all in. Per child, that is $9 to $26 a month, against $217 for the human tutor. The AI tutor is there every evening; the human is there one.

The hidden cost: the meter that rereads the whole trip

There is one catch, and it is worth understanding even if you never build anything.

A live AI conversation has no memory of its own. Every time the tutor replies, the app sends it the entire conversation so far, so it knows what was said. Fifteen minutes into a call, each reply means resending fifteen minutes of audio.

Picture a taxi meter that, at every turn, charges you again for the whole route from the start. That is what an AI voice call does, unless the provider "caches" the part it has already seen, in which case the reread costs about one percent of the normal price.

With caching working, our fifteen-minute call costs about $1.16. With caching broken, the same call costs about $11.34. Same child, same conversation, ten times the bill. For thirty children at five hours a month, that is the difference between under $800 and around $7,000.

So if you are a parent paying for a voice tutor, or a school buying one, it is fair to ask how long its calls run and whether its costs are measured from real calls. If you are building one, keep calls short, and check the first real ones before you raise anyone's limit.

Cheap is not the case

If the argument for AI tutoring were "it's cheaper," it would be a weak one. A worksheet is cheaper still.

The more important finding of the last two years is that the same technology can help a child learn or quietly stop them from learning, and the difference is in how it is built.

In a large experiment in Turkish high schools, published in PNAS in 2025, students who practiced math with an ordinary chatbot did much better while the chatbot was there, and then scored 17% worse on the exam than students who never had it (Bastani and colleagues). They had used it as a crutch. The students who used a version built to give hints instead of answers did not have that drop. Worse, the crutch group believed they had learned just as much.

That is the thing to hold onto. An AI that answers is a calculator for thinking, and a child who leans on it gets very good at leaning. An AI that tutors asks what you tried, gives one nudge, and waits. That is a design choice, made in code, and it is the only version worth paying for at any price.

The case for AI tutoring

With that caveat in place, here is why I think a well-built AI tutor is the right first tool for most families.

It shows up every night. Tutoring works mostly through repetition. The largest review of tutoring experiments found an average gain of about 0.29 standard deviations, and the strongest effects came from programs that met three or more days a week (Nickow, Oreopoulos and Quan). A human tutor at three sessions a week is $650 a month. An AI tutor at seven evenings a week is a few dollars. Frequency is the active ingredient, and AI makes frequency cheap.

It can work, when it is built to teach. At Harvard, students in a large physics course learned more than twice as much with an AI tutor designed around good teaching practice as they did in an active-learning class led by experienced instructors, in less time (Harvard Gazette). In Nigeria, a six-week after-school program using an AI tutor raised students' scores by about 0.3 standard deviations, which the World Bank estimates at close to two years of typical learning (World Bank). Both were built to guide, not to answer, and both had people around the AI, not just the AI.

It never gets tired of the same question. A twelve year old will ask an AI the same "dumb" question four times without embarrassment. Many will not ask a human even once. For a child who is quietly lost, that patience is worth more than any explanation.

It can measure what the child can do alone. This is the part I care about most. Every so often our app gives a short check with no hints and no tutor at all, and shows a parent the result. Not what the AI did, what the child did. A human tutor sees this too, but only for an hour a week.

It scales to the whole family. Two kids, three subjects, a sibling who needs help with essays and another with algebra. The human-tutor bill multiplies. The AI bill barely moves.

Where a human tutor still wins

I would not trust a post that only made one side of the case, so here is the other side, and it is real.

• The relationship. Children work for people. A tutor a child wants to impress gets effort that no app gets. Motivation is the thing AI is worst at.
• Noticing the child, not the problem. A good tutor sees the slumped shoulders, hears that it was a bad day at school, and changes the plan. An AI sees text and audio.
• Accountability. A tutor who is coming on Tuesday is a deadline. An app can be closed.
• The unusual knot. For a deep, tangled misunderstanding, or a child with a learning difference, an experienced human still diagnoses faster and better.
• The child who will not open the app. No tutor helps a child who never starts. A person at the table solves that problem in a way software does not.

Notice that most of these are about the human relationship, not about explaining math. That suggests where each one belongs.

My answer: AI for the reps, people for the relationship

So I do not think the real choice is AI or a tutor. It is how to spend the budget you have.

Think of it like fitness. A personal trainer once a week is wonderful, but the trainer is not what makes you strong. The other six days are. Most people get better results from a good plan they follow every day and a coach they see once a month than from a coach once a week and nothing in between.

For a family that would spend $217 a month on a weekly tutor, a different mix is worth considering. A good AI tutor every evening for $10 to $25, plus one human session a month for $50 to look at what the AI found, untangle what it could not, and keep the child accountable. That comes to under $80, with practice every night instead of one.

And the cheapest human tutor is already in the house. Ten minutes at dinner, asking a child to teach you what they learned, does more than most people expect. An AI tutor that tells you what to ask, and what your child can now do on their own, makes those ten minutes count.

How to tell a tutor from an answer machine

If you are choosing one for your child, five questions sort them quickly:

1. Ask it for the answer to a homework problem. Does it give it, or does it ask what your child has tried?
2. Does it ever check what your child can do with no help at all?
3. Can you, the parent, see what your child can now do, in plain words, not a chat log?
4. Who keeps your child's words, and for how long? "We don't keep them" is a good answer.
5. If it talks out loud, how long can a call run, and is there a monthly limit?

A tool that passes all five is worth a few dollars a month. A tool that fails the first one is not worth anything, however cheap it is.

The costs here are from our own spreadsheet, using list prices as of October 2026 and our app's own usage estimates. They are for planning, not an invoice, and the real numbers will come from our pilot.