Khanmigo is the AI tutor most schools should start with. It is free for teachers, and $4 a month or $44 a year for a learner or parent account (pricing, checked September 22, 2026), and it is built to walk a student toward an answer rather than hand one over. Two things change that recommendation. If you need to read every conversation a student has with the tutor, SchoolAI is the better buy. If the subject is math and the budget is a school budget rather than a family one, Third Space Learning's Skye does nothing else and does it well.
Six tools are compared below on price, teacher oversight, subject coverage and what each vendor actually says about student data. There is no single best AI tutor, because the job splits: coaching a student through a problem, watching what they ask, grading a draft, and drilling recall are four different products in practice. Our view is that most schools buy one tool for the first job and one for whichever of the other three is hurting most.
What changed in September 2026, and why it narrows the field
On September 17, 2026 the Florida Board of Education approved AI guidelines for K-12 schools and colleges (WGCU). Districts have until July 1, 2027 to write matching policies. Three of the requirements land directly on AI tutors.
Parents must be told when their child will use an AI product, consent is required, and a district that loses that consent has to offer "an alternative, non-AI instructional tool similar in instructional quality." Students may not use AI unsupervised. And districts "must prioritize vendors that keep student data within the United States."
Then there is the clause that most tutor vendors have not read yet. Florida bars AI instructional tools used "to simulate an emotional relationship with a student or use relationship-building or anthropomorphic design features to encourage a student to interact with the AI product" (ClickOrlando).
Read that next to the product category. The AI tutor market sells personality. Skye has a name and a voice. Khanmigo has a name. Aristotle's pitch is a tutor students talk to out loud. A named, chatty companion is the design pattern the whole field converged on, and it is the design pattern one large state has just written a rule against. Our reading is that the clause targets engagement-maximizing design rather than a product having a name, but nobody can say that with confidence until Florida districts start writing policy, and a Florida procurement officer reading the clause literally has a problem with most of this list. That is worth knowing before you sign a three-year contract.
The data residency line bites harder than it looks. Counting the directory on September 22, 2026: of 138 listed tools, 44 record a student-facing or tutoring use case. Of those 44, only 17 record their data hosting as the United States. Five record no hosting location at all, and 13 record no minimum user age. A district that has to prioritize US-hosted vendors is choosing from a much shorter list than the category page suggests.
Florida is one state. Idaho, Maryland, Oklahoma and Virginia moved on district AI policy in the same cycle, and the FutureEd legislative tracker is the place to check your own. The direction is consistent enough to plan around: notify, consent, supervise, and know where the data sits.
Three questions that pick your AI tutor
Work through these in order. They eliminate faster than a feature comparison.
1. Who is buying, the school or the family? This splits the market cleanly and almost nobody says so. Khanmigo, Knowt and Aristotle sell to individuals as well as institutions. Skye sells only to schools: "This service is currently only available for schools," its pricing page states. SchoolAI and Class Companion are school purchases with a free entry point for a single teacher. If you are a parent, half this list is closed to you. If you are a district, the family-priced tools arrive without a data agreement and you will need one.
2. Does a teacher need to see the conversation? Some tutors are a private space between a student and a model. Others put every exchange in front of the teacher. That is not a feature preference; under Florida's supervision rule it is close to a requirement, and it is the single biggest difference between the tools below. SchoolAI is built around it. Khanmigo gives teachers activity summaries rather than a live feed.
3. Is the subject math, writing, or everything? General tutors are weaker at both ends than specialists. A math-only tool that models working out beats a general chatbot at algebra, and a writing tool that reads a full draft beats both at an essay. If one subject is the reason you are shopping, buy for that subject.
The six compared
| Tool | Best use | Strengths | Limitations | Price / free tier | Privacy considerations |
|---|---|---|---|---|---|
| Khanmigo | A school's first AI tutor, math-leaning | Socratic prompting, wide subject range, published price | Teacher sees summaries, not live chats | Free for teachers; $4/mo or $44/yr learners and parents; district pricing on request | Khan Academy privacy policy; directory record blank on FERPA, COPPA, hosting and age |
| SchoolAI | Any district that must supervise student AI use | Teacher reviews every student conversation | No published price; free trial capped at 5 Space launches a year | Free trial; Pro and Scale by demo | Vendor states FERPA, COPPA and SOC 2 Type 2 compliance and 1EdTech certification; directory records US hosting and third-party data sharing |
| Bloom AI | Tutoring grounded in your own course materials | Tutor answers from your documents; published per-student pricing | Analytics capped at 7 days on the free plan | Free ($0, up to 3 staff and 30 students per Teaching Space); Educator Plus $12/mo billed annually; Growth from $120/mo | Vendor states alignment with FERPA, COPPA, GDPR, HECVAT and Australian Privacy Principles; directory records hosting as predominantly Australia |
| Skye (Third Space Learning) | Math intervention bought by a school | Built for one subject; regional data hosting | Math only; UK pricing and curriculum framing | £3,500 to £6,000 a year by school size; free trial offered | Directory records UK or EU hosting for UK student data and US hosting for US student data; privacy policy |
| Class Companion | Coaching a student through a draft | Feedback on work in progress, free plan for one teacher | Directory record is unreliable (see Safety) | Free plan for individual teachers and their students; school and district plans on request | Vendor states FERPA, COPPA and state privacy law compliance, Student Privacy Pledge signatory, TrustEd Apps certified |
| Knowt | Retrieval practice and revision on a zero budget | Genuinely useful free tier; incident monitoring included | Drill and recall, not conceptual coaching | Free Basic plan; Ultra $199.99 a year ($16.99/mo equivalent); school site pricing by volume | Privacy policy; directory records US hosting and FERPA, COPPA and GDPR as yes |
Prices were read from each vendor's own pricing page on September 22, 2026.
Khanmigo: the best fit for a school's first AI tutor
Khanmigo is the best fit for a school taking its first step into AI tutoring, because it is the only tool here that costs a teacher nothing and still publishes a price for everyone else. Khan Academy describes it as guiding learners to find answers themselves rather than supplying them, across math, humanities and coding. It carries a 4.40 rating from 5 reviews in our directory, the largest review count on this list after SchoolAI.
The pedagogy is the reason to pick it. A student stuck on a quadratic gets a question back, not a solution, and the model holds that line more consistently than a general chatbot asked to behave like a tutor.
Watch out for: the teacher view is activity summaries and student work, not a live feed of every exchange. If your district's policy requires supervision of each conversation, that gap matters and Khanmigo alone will not close it.
Pricing: free for teachers, with lesson planning, student work summaries and rubrics included. Learners and parents pay $4 a month or $44 a year. Districts request pricing, which adds SSO, analytics and professional development (pricing, checked September 22, 2026).
Safety: the pricing page states no age minimum, and Khan Academy's privacy policy governs the service. Our directory record for Khanmigo is blank on FERPA, COPPA, hosting location and minimum age, which is a gap in our own data rather than a finding about the product. A district will need to request the data privacy agreement directly and confirm where student data sits before a rollout, particularly under a state rule that prioritizes US hosting.
SchoolAI: the best fit for a district that must supervise every conversation
SchoolAI is the best fit for a district that has to supervise student AI use, because supervision is the product rather than a setting inside it. Its pricing page lists "Student Chat Observability: Review and monitor all student AI conversations" on the Pro and Scale tiers. Teachers build a Space, set what the tutor will and will not do, and read what students actually asked. It holds a 4.92 rating from 13 reviews in our directory, the highest count of any tool in this category.
That last part is worth more than it sounds. A teacher reading twenty transcripts on Friday learns which misconception is general and which is one student's, and that is diagnostic information no quiz returns.
Watch out for: no published price, so a comparison against Khanmigo's $44 a year is impossible until you have had the demo. The free trial is capped at 5 Space launches a year, which is enough to evaluate and not enough to teach with.
Pricing: Free Trial, Pro and Scale, all quoted by demo. The free trial includes 5 Space launches a year, limited browser extension availability and access to the premade Spaces library (pricing, checked September 22, 2026).
Safety: the vendor states FERPA, COPPA and SOC 2 Type 2 compliance and 1EdTech certification, with a standard data privacy agreement on the Free and Pro tiers and custom redlines available on Scale. Our directory record lists US data hosting and records that the tool shares data with third parties, which is the line to raise in your own review. No age minimum is stated, so the age floor is whatever your district sets.
Bloom AI: the best fit for tutoring that must stay inside your own materials
Bloom AI is the best fit when the tutor has to answer from your curriculum rather than from the open internet, because the product is built around grounding a tutor on course materials you upload, with what the vendor calls pedagogical guardrails. For a department that has spent years writing its own schemes of work, a tutor that will not wander outside them is the whole proposition.
It is also the only tool here that publishes a per-student price, which makes budgeting honest: $120 a month billed annually covers up to 30 students, and the tiers run to $2,000 a month for up to 500.
Watch out for: the free plan caps analytics at 7 days and limits advanced model messages to 50 a week and Deep Think messages to 5 a week. Those limits are fine for evaluating and thin for a term.
Pricing: Free at $0 with no card required, covering up to 3 staff and 30 students per Teaching Space. Educator Plus is $12 a month billed annually. Growth runs from $120 to $2,000 a month billed annually by student count. Enterprise is custom (pricing, checked September 22, 2026).
Safety: the vendor states alignment with FERPA, COPPA, GDPR, HECVAT and the Australian Privacy Principles on all plans, and our directory records a minimum age of 13 and hosting as predominantly Australia with other regions available. A US district working to a US-hosting preference should confirm which region its own instance runs in before signing, and that is a question the pricing page does not answer.
Skye from Third Space Learning: the best fit for math intervention on a school budget
Skye, the AI math tutor from Third Space Learning, is the best fit for scheduled math intervention, because it is sold as a school service with a fixed annual price and no family tier at all. Third Space Learning describes it as built by math teachers, adaptive and aligned to your curriculum or standards, with unlimited sessions.
A fixed annual price is unusual in this market and it solves a real problem. A school can put Skye in front of every child in a year group without a finance conversation about seat counts.
Watch out for: math and nothing else, and the framing is British. The pricing page prices by "one-form entry" and "Year 3-6," which a US reader has to translate. There is a US version of the site, but the published prices are in pounds.
Pricing: £3,500 a year for a one-form-entry primary school, £5,000 for two-form, £6,000 for three-form, and £5,000 for any secondary school covering Year 7 and GCSE. A free trial is offered without a stated duration (pricing, checked September 22, 2026).
Safety: our directory records a minimum age of 8, FERPA, COPPA and GDPR as yes, no third-party data sharing, and hosting split by region, with UK student data in the UK or EU and US student data in the US. That regional split is the most useful thing on this list for a district working to a data residency rule. It comes from our own directory record: Third Space Learning publishes its data protection and privacy policy as an embedded PDF rather than as text, so treat the clause as unverified until you have it in writing from the vendor.
Class Companion: the best fit for coaching a student through a draft
Class Companion is the best fit for the gap between a student writing something and a teacher reading it, because it tutors on work in progress and shows the teacher the trace. The vendor lists English, science, social studies and world languages, kindergarten through higher education, plus test preparation for CAASPP, STAAR and AP. It carries a 4.00 rating from 1 review in our directory.
Watch out for: our own record for this tool is contaminated and you should not rely on it. It carries another vendor's plan names and another vendor's privacy policy URL, and its compliance and hosting fields all read "Unknown" while the vendor's own site states them plainly. We have flagged it for correction. Use classcompanion.com as the source until it is fixed.
Pricing: the vendor states a free plan for individual teachers and their students, with paid school and district plans adding LMS integration, insights and response integrity detection. No public per-seat price (classcompanion.com, checked September 22, 2026).
Safety: the vendor states compliance with FERPA, COPPA and state privacy laws, that it is a Student Privacy Pledge signatory and TrustEd Apps certified for data privacy, and that it does not market, sell or rent user personal information or use data for advertising. It states no minimum age. Those are the vendor's own statements and they are more specific than most; your district still signs its own agreement and confirms the age floor for the grades you intend to use it with.
Knowt: the best fit for retrieval practice when the budget is zero
Knowt is the best fit for revision and recall on no budget, because its free plan gives unlimited AI use for teachers and students on teacher-shared files rather than metering it. It holds a 5.00 rating from 1 review in our directory. Flashcards, notes, practice tests and classroom games are the shape of it.
Call it what it is. This is drill, and drill is the part of tutoring that scales cheapest and matters more than its reputation suggests. A student who cannot recall the formula cannot reason with it.
Watch out for: it will not coach a student through a misconception the way Khanmigo or Bloom will. Buying Knowt as your conceptual tutor is buying the wrong tool. The free plan caps file creation and storage at 20.
Pricing: Basic is free and includes unlimited AI usage for teachers and students on teacher-shared files, incident monitoring, and 20 file creations. Ultra is $199.99 billed upfront, which the vendor shows as $16.99 a month. School site licenses are priced by volume (teacher plans, checked September 22, 2026).
Safety: our directory records FERPA, COPPA and GDPR as yes, US hosting, no third-party sharing, and deletion on request, with the vendor's privacy policy as the source. Knowt records no minimum user age, so an elementary rollout needs a COPPA conversation with your district before students create accounts. Our guide to AI tools for elementary students covers what changes under 13.
The big names people ask about, and why they are not on this list
Search engines and AI assistants keep offering these four when someone asks for an AI tutor. Each does something real. None is a tutor in the sense used above.
ChatGPT is a general assistant, not an education product. It has no teacher dashboard, no curriculum grounding and no district data agreement by default. Under a state rule requiring supervised use and a non-AI alternative, a consumer chatbot account is the hardest thing in a school to defend. Used by a teacher to build materials it is excellent; handed to a student as a tutor it is unsupervised by design.
Gemini Notebook, the tool Google renamed from NotebookLM in July 2026, grounds answers in documents you give it, which makes it the closest thing on this list to Bloom's approach without the classroom layer. Good for a student revising from their own notes. No oversight, no age controls, no reporting.
Quizlet is a study tool with AI features bolted on rather than an AI tutor. If retrieval practice is the job, compare it against Knowt on price and on what each vendor says about student data.
MagicSchool and Diffit are teacher-facing generators. They build the worksheet, the text at the right reading level, the questions. That is a different job from sitting with a student while they work, and our worksheet generator comparison is the better page for it.
What the research actually shows
The strongest evidence available is a randomized controlled trial published in Scientific Reports on June 3, 2025 (Kestin et al.). One hundred and ninety-four Harvard undergraduates in an introductory physics course were randomly assigned, in a crossover design, to a purpose-built AI tutor or to in-class active learning. Median learning gains in the AI-tutored group were more than double those of the in-class group, with an effect size of 0.63 standard deviations in the linear model, and students reported higher engagement and motivation. Median time on the AI tutor was 49 minutes against roughly 60 minutes in class.
Read the limitations the authors state, because they are the part that applies to your school. The lessons targeted the understanding, applying and analyzing levels of Bloom's taxonomy, and the authors do not claim the result extends to complex synthesis. The design depended on expert-crafted, question-specific prompts and high-quality instructional videos, which is to say on a great deal of teacher work done in advance. The study covered two lessons, not a course. And the participants were university students in one subject at one institution.
So the honest summary is narrower than the headline. A carefully built AI tutor, pointed at well-understood content, with a teacher's thinking baked into the prompts, outperformed a good classroom lesson for two topics with undergraduates. That is a real finding and a genuine reason to try this. It is not evidence that handing a twelve-year-old a chatbot raises attainment, and no study currently shows that.
A worked example: one unit, one tutor, one consent form
Here is the shape of a first deployment that survives a policy review. Seventh grade math, a three-week unit on ratio and proportional relationships, roughly 60 students across two classes.
Before the unit. The teacher picks one tool, Khanmigo, because it costs the school nothing and the district already has Khan Academy. The office sends families a notice naming the product, what it will be used for and for how long, with a reply slip for consent. For any family that declines, the alternative is the same practice set on paper with the teacher's own worked examples, which the unit already contains. That alternative is drafted before the notice goes out, not after the first refusal arrives.
During the unit. Tutor use happens in class, in the last fifteen minutes of three lessons a week, on school devices, with the teacher circulating. Nothing goes home. That single decision satisfies a supervision requirement, removes the device-equity problem and keeps the whole thing inside a window the teacher can observe.
What the teacher watches. Not scores. The question is which step students get stuck on, and the answer arrives as a pattern: most of them can scale a ratio up and stall when asked to scale one down. That is a reteach for Thursday, and it is the output that justifies the tool.
After the unit. One measure, chosen in advance: the end-of-unit assessment against the same assessment from last year's cohort. Not a controlled comparison and nobody should pretend otherwise, but it is enough to decide whether to run it again with eighth grade.
The pattern generalizes. One tool, one unit, one class set, one measure, one alternative ready for the families who say no. Browse the full tutoring and personalised learning category for the current field, and see our guide for tutors working professionally if you are running sessions rather than classes.
Where AI tutoring still does not belong
Keep four things human. Judging whether a student's confusion is conceptual or emotional. Deciding what a struggling reader needs next. Any conversation about effort, identity or whether someone is capable. And assessment that counts.
The Florida rule against tools designed to meet a student's social or emotional needs is clumsy in places, but it is pointed at something real. A tutor that is pleasant to talk to is useful. A tutor a lonely child prefers to a person is a different product with a different set of risks, and the field has not yet drawn that line for itself.
For students who need more than pacing support, our guides to AI tools for special education teachers and to AI tools for gifted students cover the two ends of that range, and AI tools for middle school teachers covers the grades where most of these tools land first.
