AMP vs ChatGPT

ChatGPT, Claude and Gemini will explain almost any math problem at any hour, and each has a free plan2,7,10. We found no sequence of topics, mastery rule or review schedule in their makers' descriptions of their learning modes1,9,13, and all three makers warn that answers can be wrong5,8,12. AMP chooses every problem, checks every answer with tested code, and schedules review, with no AI model running while a child works.

Try it free

Why AMP

AMP decides what comes next

OpenAI describes study mode as a way to work through problems step by step, with questions that assess skill level and memory from earlier chats1. That description has no placement test, skill map or sequence of topics1. AMP keeps a level in each of 14 strands of math, places the learner within the first 24 problems, and chooses every next problem from the learner's answers: new work, a prerequisite, or review. With a chatbot, the adult has to plan all of that.

The research: Practice at the learner’s level

Help in AMP cannot turn into the answer

In a randomized field experiment with nearly 1,000 high school math students, unrestricted GPT-4 raised practice scores by 48%, and once access was removed those students scored 17% lower on exams than students who never had it18. A guarded tutor version largely mitigated that harm18. ChatGPT's study mode also aims to guide instead of answer, and OpenAI lists toggling it on and off during a conversation as a feature1. A parent with a linked teen account can turn it on from parental controls, so each new chat starts in it; OpenAI's page does not say it can be locked on6. AMP's hint gives one step and never the answer, and an answer given after a hint does not count toward mastery.

The research: One-step hints · The research: Mastery, and repair of gaps

Every answer is checked by code, not a model

OpenAI, Anthropic and Google each warn in their terms or help pages that output can be wrong5,8,12. Anthropic's terms add that users should not rely on outputs without confirming their accuracy8, and OpenAI says study mode may make mistakes across conversations1. In AMP no AI model runs while a learner works: every problem, hint and worked solution comes from tested code, and 185 independent checkers re-derive the answers.

The research: Feedback after every answer

Learned skills come back on a schedule

Study mode and Guided Learning quiz the learner inside a chat1,13, and ChatGPT can remember earlier chats2. Their makers' descriptions, and Anthropic's description of Claude's learning mode, give no rule for when a skill is learned and no schedule for bringing it back1,13,9. AMP marks a skill Strong at four of the last five answers right without help, then brings it back after 6 hours, 1 day, 3 days, 1 week, 2 weeks and a month, mixed with other topics. In a randomized trial with 787 seventh graders, mixed and spaced practice scored 62% on the researchers' test a month later, against 38% for practice in blocks22.

The research: Mastery, and repair of gaps · The research: Spaced, mixed review

Built for children from kindergarten up

OpenAI's terms require ChatGPT users to be at least 13, with a parent's permission under 185. Claude's consumer terms require users to be at least 188. Gemini is for 13 and over11; a younger child needs a parent to turn it on through Family Link, and Google tells parents its filters are not perfect12. In AMP a child is never an account holder: an adult adds the child, AMP stores only a birth year, and the problems start at counting.

The adult sees what was learned

ChatGPT's parental controls let a parent set quiet hours, turn off memory or voice, and get alerts on signs of acute distress6. They are safety settings. Study mode's quizzes give the learner feedback inside the chat1, and OpenAI's descriptions of parental controls and study mode include no progress report for the adult and no skill map6,1. AMP shows the level in every strand with its likely range, the skills earned, and time by day, and the record opens to the actual problems with the help used. It also prints workbooks made for the learner, each with a worked answer key, and results entered from paper count.

The research: Progress parents and teachers can see

Side by side

AMPChatGPT
Price$599 a year or $59 a month for the first child, then $399 or $39 for each more. Two parents practice free. Full refund within 30 days.Free plan. Go $8 a month in the U.S., which may include ads; Plus $20 a month, billed monthly2,3,4. Claude Pro $20 billed monthly7. Google AI Plus $4.99 a month10.
Who can buyFamilies, tutors and schools, directly. Free to start without a card: 100 answers and 5 workbooks. Arizona ESA, Arkansas EFA and New Hampshire EFA funds pay for it.Anyone old enough. ChatGPT: 13 and up, with a parent's permission under 185. Claude: 18 and up8. Gemini under 13 only through a parent12.
RangeCounting through precalculus in one program: 336 skills on one map.A general chatbot: no grade range. The learner brings the topic; study mode's examples include Bayes' theorem and game theory1.
Finding the starting pointNo separate test. Placement takes the first 24 problems of practice. After that, every answer adjusts a level in each of 14 strands of math.No placement test. Study mode assesses skill level with questions and memory from earlier chats1. OpenAI describes no stored or shown level.
TeachingA one-step hint and a written lesson with a worked example for every problem type. Worked steps after every answer. Narrated lesson films.Explanations in text or voice on request2. Study mode asks guiding questions, gives hints and quizzes1. OpenAI warns output may not always be accurate5.
What decides the next problemAMP chooses every problem from the learner’s answers: new work just within reach, a prerequisite, or review.Whatever the learner or the adult asks next. OpenAI's description of study mode sets no sequence of topics1.
MasteryFour of the last five answers right without help. An answer after a hint doesn’t count. Two misses in a row send practice back to what the skill builds on.No mastery rule found in the published descriptions of study mode, learning mode or Guided Learning1,9,13.
ReviewScheduled. Learned skills come back after six hours, a day, three days, a week, two weeks and a month, mixed with other topics.No scheduled review found in the makers' descriptions1,9,13. ChatGPT can remember earlier chats2.
PrintUnlimited workbooks made for each learner, each with a worked answer key. Results entered from paper count.A chat app, not a workbook program. Any worksheet it writes falls under OpenAI's warning that output may not always be accurate5.
What the adult has to doNothing to plan, teach or grade. AMP places the learner, chooses each problem, checks it and schedules review.Choose the topics and their order, find the problems, judge when a skill is learned, plan review, keep records, and check that study mode stays on1.
ReportsThe level in every strand with its likely range, skills, time, and the actual problems behind each number. For teachers, who needs them today.The chat history. ChatGPT's parental controls cover quiet hours, memory, voice and distress alerts, not learning6.
EvidenceA published rationale (ESSA Tier 4) with the research behind each part, public source code, and outcome studies from 2027, each registered before it starts.No outcome study of study mode, learning mode or Guided Learning found; OpenAI says research is still taking shape1. No entry in Evidence for ESSA or the What Works Clearinghouse20,21.

What ChatGPT is

From OpenAI

This page is about using a general AI chatbot as a child's math tutor in place of a math program. ChatGPT, from OpenAI, is the best known: OpenAI reported 700 million weekly active users in September 202516, and 26% of U.S. teens said in fall 2024 that they had used it for schoolwork17. The same points apply to Anthropic's Claude and Google's Gemini.

A session has no set shape. A learner types or photographs a problem, often from another book or a homework sheet, and the chatbot explains, by text or by voice2. OpenAI added study mode in July 2025 so that students work through problems step by step with guiding questions, hints and quizzes instead of only getting an answer1. Google's Guided Learning for Gemini does much the same13, and Anthropic built a learning mode for Claude for Education9.

None of the three is sold as a math curriculum. The learner or the adult decides what to ask, in what order, and when to move on1.

What ChatGPT does well

The evidence

ChatGPT

We found no study of ChatGPT's study mode, Claude's learning mode or Gemini's Guided Learning that measures K-12 math outcomes. OpenAI writes that research on AI in education is still taking shape and that longer studies will follow1. Google's Guided Learning announcement reports no outcome data13, and the LearnLM technical report evaluates the model family behind it by expert raters' preferences, not by what students learned14. General-purpose AI chatbots have no entry in Evidence for ESSA or the What Works Clearinghouse, as of November 202620,21.

The most direct test of a chatbot as a math helper is a field experiment by Bastani and colleagues, published in PNAS in 202518. Classes of nearly 1,000 students in grades 9 to 11 at a Turkish high school were randomly assigned to plain GPT-4, a GPT-4 tutor given teacher-written solutions and told to hint without revealing answers, or no AI, over four 90-minute sessions. Plain GPT-4 raised practice scores by 48% and lowered later exam scores, taken without AI, by 17%. The tutor version largely mitigated that harm; its exam scores were statistically indistinguishable from the control group's18.

Khanmigo is Khan Academy's AI tutor, a different product, but its trial shows what happens when a guarded chat tutor is offered every day. In a two-year randomized trial in 18 Tennessee middle schools, adding Khan Academy with Khanmigo to daily remedial math raised scores by about 0.06 to 0.08 standard deviations a year, gains the authors say resemble what Khan Academy practice without AI predicts. The median student messaged Khanmigo in only 17% of the exercise sessions in which they made a mistake19. The paper is a working paper, not yet peer reviewed.

AMP

Two rules in AMP answer what these trials found. Help never replaces the work: a hint gives one step and never the answer, and a hinted answer does not move the level. And help does not wait to be asked: after every answer AMP shows the correct answer and the worked steps. Across 40 studies of computer-based feedback, feedback that explains averaged an effect of 0.49, against 0.32 for the correct answer alone and 0.05 for right or wrong only23. We think a set path, exact checking and scheduled review will do more for most children than an open chat window, and we have committed in public to measuring AMP's results.

AMP is new, and its own outcome studies have not run yet. They begin in 2027: growth on a national test against hours of use, then comparison groups, then randomized trials. Each will be registered before it starts and published whatever it finds. Until then, the research behind each part of AMP is laid out in our report, and the code is public.

The research behind AMP

Choose ChatGPT if

Choose AMP if

Questions

Can ChatGPT replace a math curriculum for my child?

OpenAI does not offer it as one. Study mode helps a student work through problems step by step with questions, hints and quizzes1, but OpenAI's description sets no sequence of topics, mastery rule or review schedule1. The adult has to supply those, and that is the work AMP does.

Does using ChatGPT for math hurt learning?

It can, depending on how it is used. In a randomized field experiment, high school students with unrestricted GPT-4 scored 17% lower on later exams than students without it, and a guarded tutor version largely mitigated that harm18. Study mode is meant to guide instead of answer, and a student can toggle it off during a conversation1.

How old does a child have to be to use ChatGPT, Claude or Gemini?

ChatGPT: at least 13, with a parent's permission under 185. Claude: at least 188. Gemini: 13 and over, or younger if a parent turns it on through Family Link11,12.

Is AMP more expensive than ChatGPT?

Yes. AMP is $599 a year or $59 a month for the first child and $399 a year or $39 a month for each additional child, while ChatGPT has a free plan, Go at $8 a month and Plus at $20 a month, billed monthly2,3,4. AMP's price buys the plan a chatbot leaves to the adult: placement, every next problem, exact checking, scheduled review, and printable workbooks with answer keys, from counting through precalculus. The first 100 answers are free, and there is a full refund within 30 days.

Does AMP use AI?

Not while a learner works. Every problem, hint and worked solution comes from fixed code that is tested and checked exactly, and 185 independent checkers re-derive the answers. The code was written with the help of AI coding tools, and it is public for anyone to read.

Sources

  1. Introducing study mode (OpenAI, July 2025)
  2. ChatGPT pricing (OpenAI)
  3. What is ChatGPT Plus? (OpenAI Help Center)
  4. Introducing ChatGPT Go (OpenAI, January 2026)
  5. Terms of Use (OpenAI)
  6. Introducing parental controls (OpenAI)
  7. Claude plans and pricing (Anthropic)
  8. Consumer Terms of Service (Anthropic)
  9. Introducing Claude for Education (Anthropic, April 2025)
  10. Google AI plans (Gemini)
  11. Gemini Apps Help: What you need to sign in to Gemini Apps (Google)
  12. Gemini Apps Help: Guide your child's Gemini Apps experience (Google)
  13. Guided Learning in Gemini (Google, August 2025)
  14. LearnLM Team (2024), LearnLM: Improving Gemini for Learning (arXiv)
  15. Gemini Apps Help: Where you can use the Gemini web app (Google)
  16. How people are using ChatGPT (OpenAI, September 2025)
  17. About a quarter of U.S. teens have used ChatGPT for schoolwork (Pew Research Center, January 2025)
  18. Bastani et al. (2025), Generative AI without guardrails can harm learning (PNAS)
  19. Oreopoulos & Low (2026), One click away: AI tutoring with Khanmigo (NBER w35620)
  20. Math programs (Evidence for ESSA)
  21. Intervention reports (What Works Clearinghouse)
  22. Rohrer, Dedrick, Hartwig & Cheung (2020), A randomized controlled trial of interleaved mathematics practice (What Works Clearinghouse review)
  23. Van der Kleij, Feskens & Eggen (2015), Effects of feedback in a computer-based learning environment, Review of Educational Research 85(4)

Checked November 2026.

ChatGPT is a trademark of its owner. AMP is not affiliated with it. Found an error? Write to skylar@corpora.inc.

More comparisons

All comparisons · AMP · Research