18 AI Tools for Education, Fact-Checked for 2026
This post used to recommend 25 tools. Seven of them are gone from this update, and three of those seven no longer exist at all: Tome serves a Vercel error page, Knowji’s domain is parked at GoDaddy, and ScribeSense returns HTTP 200 over a domain-parking stub with no product behind it.
The pricing was worse. Checking every surviving tool against its own pricing page in July 2026, nearly every number in the old version was wrong. Curipod was listed at $9.99 a month for a “Pro” tier that does not exist; the real plan is Premium at $24 billed monthly. Synthesis Tutor was listed at $149 a month, a figure that appears nowhere on Synthesis’s site. Khanmigo was described as “3.0, powered by GPT-5” when Khan Academy has never published a version number and its own page says GPT-4.
That is the normal condition of AI tool roundups, including the ones ranking above this one. So here are 18 tools that survived a check against primary sources, what each actually costs this month, and the three questions that matter more than any feature list: who is allowed to sign the contract, what happens to student data, and whether the thing works.
Read this before you sign up for anything
The tier you buy is a compliance decision, not a budget decision. This is the part every roundup skips, and it is the part that gets schools in trouble.
Under FERPA, a vendor can touch student records through the “school official” exception in 34 CFR 99.31, but only if it performs a service the school would otherwise use employees for and stays under the direct control of the school. A consumer terms-of-service that lets the vendor use your inputs to improve its models is the opposite of direct control. The Department of Education’s Privacy Technical Assistance Center is blunt about the click-wrap problem: accepting terms “serves to enter the provider and the end-user (in this case, the school or district) into a contractual relationship akin to signing a contract.” A teacher clicking accept can bind the district to terms nobody reviewed.
COPPA tightened in the meantime. The amended Rule was published at 90 FR 16918, took effect 23 June 2025, and the general compliance date of 22 April 2026 has now passed, so there is no grace period left. Biometric identifiers such as voiceprints and facial templates are now personal information, which is what bites AI proctoring. Retention policies must be publicly posted, meaning you can read a vendor’s before you sign. Notably, the FTC declined to codify the school-consent mechanism, so it still rests on staff guidance rather than regulation. That guidance names a red flag worth memorising: operators “should not state in Terms of Service or anywhere else that the school is responsible for complying with COPPA.” If a vendor’s contract pushes COPPA onto you, the FTC has already said that is backwards.
EU institutions have a sharper problem. Legitimate interests is unavailable to public authorities under Article 6(1), and the EDPB considers consent presumptively invalid in schools because of the power imbalance. Article 28 requires the vendor to process data “only on documented instructions,” which a vendor training on your inputs is not doing. And the SCHUFA ruling (C-634/21) closed the obvious escape hatch: an automated score still falls under Article 22 if the human downstream gives it a determining role. “The teacher makes the final call” does not save an AI grading flag if the teacher rubber-stamps it.
The practical version fits in four checks. Does the district approve the tool, not the individual teacher? Can the vendor use inputs for training? Is there a published retention policy with a deletion window? And is there a signed agreement at all, as opposed to a personal card on a personal seat?
That last one decides more than people expect. OpenAI, for instance, states that it can execute a Data Processing Addendum for ChatGPT Business, Enterprise and the API, and handles Edu and Teachers under a separate Student Data Privacy Agreement. Free and Plus appear nowhere on that list. The agreement is what makes OpenAI “a School Official… under the direct control and supervision of Customer” in writing, and it attaches to a signed order form. An expensed personal seat has no such document.
AI tutoring and adaptive courseware
1. Khanmigo
Free for teachers. That is the headline, and Khan Academy has kept it true.
Khanmigo tutors with a Socratic approach, refusing to hand over answers, which is either the whole point or deeply annoying depending on the student. It covers Khan Academy’s K-12 curriculum, gives teachers a lesson-planning side, and reports to parents. Family access runs $4 a month or $44 a year. Districts get a quote rather than a price.
Two corrections to what you will read elsewhere. There is no “Khanmigo 3.0” and no “Khanmigo Plus” tier; both are inventions that spread through AI-written listicles. And it is not running GPT-5. Khan Academy’s own parent page says GPT-4, and its engineering blog describes GPT-4 Turbo with GPT-4o in testing. That copy may itself be stale, which is the honest position: Khan Academy has not published a current model disclosure, so anyone quoting a specific model is guessing.
Weakest at specialised or non-US curricula, where the coverage thins out fast. khanmigo.ai
2. Synthesis Tutor
A K-5 maths tutor for ages 5 to 11, grown out of the programme originally built for SpaceX employees’ children. It is good at the thing most maths software is bad at: making a child reason rather than pattern-match to the answer box.
Pricing is where this gets awkward. Individual access is $45 a month billed monthly, or $29 a month billed annually, with a $999 lifetime option. Family plans for up to 10 children run $70 monthly or about $33 a month annually. But Synthesis’s own page serves several conflicting price tables, apparently from unfinished A/B tests, and in mid-July it was still advertising a holiday sale and classes beginning 17 February. Treat any number, including these, as approximate until checkout. A pricing page that contradicts itself is itself a signal about how closely the product is being tended.
Narrow by design: it is maths, for young children, as a supplement. It will not carry a curriculum. synthesis.com/tutor
3. Knewton Alta (now a Wiley product)
Alta is adaptive courseware for college STEM: maths, chemistry, physics, statistics, economics. Students who cannot do a problem get walked back to the prerequisite they actually missed, which is the one trick textbooks cannot do.
knewton.com now redirects through four hops to Wiley, and Alta sits in Wiley’s navigation next to WileyPLUS and zyBooks with no sunset notice anywhere. Worth saying plainly, because the assumption runs both ways: Alta was not retired in the Wiley acquisition. It is alive and still sold. Pricing is $49.95 for single-term access, $99.95 for an AltaPass covering every course in a discipline for two years, or $14.95 a month. Institutional deals are negotiated separately.
Restricted to the subjects Wiley has built, and you are buying into a publisher’s ecosystem. Knewton Alta at Wiley
Tools that cut teacher workload
4. MagicSchool AI
The default choice for teacher-facing AI, and the free tier is genuinely free rather than a trial. It bundles 80-plus teacher tools (lesson plans, rubrics, IEP drafting, report-card comments, parent emails with translation) plus a separate MagicStudent side with 50-plus student tools that most write-ups omit.
Plus is $12.99 a month billed monthly, or $8.33 per user per month billed annually. It is not $14.99; that figure does not appear on the pricing page.
MagicSchool advertises SOC 2, FERPA and COPPA compliance and a 1EdTech certification. Those are vendor self-claims and no audit report is published, which is not an accusation, only the correct way to read a trust page. Output quality swings hard by tool: the lesson planners are solid, the comment writers produce filler you will rewrite. If it is not clicking for you, we compared the field in MagicSchool AI alternatives. magicschool.ai
5. Curipod
Type a topic, get an interactive lesson with polls, word clouds and open-response prompts that students join on their devices. For getting a room to participate, it earns its place.
The price is the story. Premium is $24 a month billed monthly, or $228 a year, and the page defaults to the yearly toggle, which is how the old $9.99 figure survived in so many roundups. There is no Pro tier. Curipod AS is Norwegian and states it is FERPA, COPPA and GDPR compliant with student data privacy built in, which is a stronger starting position than most US-only vendors for European schools.
One caution: their pricing page carries visible locale bugs, including a stray $25 monthly string and a line mixing dollars with pounds. Check what your checkout actually says. Needs a projector and live internet, so it is useless as homework. For the video-lesson equivalent, see Edpuzzle alternatives. curipod.com
6. Eduaide.AI
The cheapest serious tool here. Eduaide Pro is $5.99 a month billed monthly, or $49.99 a year, which works out near $4.17 a month. The free tier is permanent with generation limits.
It generates worksheets, rubrics, discussion questions and feedback, and aligns to standards from 50-plus jurisdictions. It overlaps heavily with MagicSchool at a third of the price, and for a teacher paying out of pocket that maths is hard to argue with.
The commonly repeated “100+ resource types” claim could not be verified against anything Eduaide publishes, so treat it as marketing that escaped into the wild. eduaide.ai
7. Diffit
Diffit does one job. It takes a text and rewrites it at several reading levels while keeping the content intact, then generates vocabulary and comprehension questions for each.
For a mixed-ability class or an English-language-learner cohort, this replaces an evening of work. Individual teacher access is $14.99 a month or $149.99 a year, credit card only, and individuals cannot pay by purchase order. School licences are flat-rate annual, tiered by enrolment, and those do take POs. Diffit states it meets FERPA and COPPA standards.
Built for informational text. Feed it poetry or fiction and the adaptation flattens exactly what you wanted students to notice. diffit.me
Assessment, grading, and the detection mess
8. Gradescope
Gradescope is a Turnitin product now, and it remains the best answer to grading handwritten maths and problem sets at scale. Scan the stack, group identical answers, apply a rubric once, propagate it.
The free Basic plan still exists, contrary to a rumour worth putting down: Turnitin says outright that after an institutional trial ends “you can continue using the Basic plan for free.” Basic gives you dynamic rubrics, question-by-question grading, statistics, regrade requests and data export.
Here is the catch nobody states clearly. The AI grading is not in the free tier. AI-assisted grading, answer groups, multi-grader workflows, anonymous grading, LMS integration, SSO and code similarity are all gated behind the Institutional licence, which is quote-only. If you adopted Gradescope expecting the AI features on a free instructor account, that is the disappointment waiting for you. Gradescope pricing
9. Turnitin
Turnitin’s similarity checking is still the sector standard and is not really in dispute. Its AI writing detection, sold as part of Originality, very much is.
To Turnitin’s credit, it publishes its own caveats, and they are sharper than most institutions realise. It reports a document false-positive rate under 1%, but only for documents already scoring above 20% AI writing. Below that threshold the company concedes “a higher incidence of false positives” and now marks those scores with an asterisk. It raised the minimum document length from 150 to 300 words. Its sentence-level false-positive rate is roughly 4%, meaning about four highlighted sentences in every hundred were written by a human. And it says plainly that “Turnitin does not make a determination of misconduct.”
Read that back and the shape becomes clear: for sub-20% documents, no false-positive rate has ever been published. Institutional licensing, quote-only. turnitin.com
What the research actually says about AI detectors
The study everyone cites is Liang et al., “GPT detectors are biased against non-native English writers” (Patterns, July 2023). Seven detectors misclassified 61.3% of TOEFL essays by Chinese students as AI-generated, against 5.19% for US eighth-grade essays. You will see 97% quoted; that is the share flagged by at least one detector and it overstates the finding roughly fivefold. The mechanism is the damning part: rewriting the American essays with richer vocabulary, changing nothing about their authorship, pushed misclassification from 5.19% to 56.65%. The detectors were measuring vocabulary, not authorship. Two caveats matter for fairness: those were consumer-grade GPT-3.5-era detectors, and Turnitin was not among them.
Turnitin’s own research claims no statistically significant bias against English-language learners, and that claim is not nothing. Read its caveat, though, and vendor and critics converge: the finding holds only above 300 words, and Turnitin concedes the gap between groups widens significantly below that. Short work by a non-native writer is the danger zone, and both sides agree on it. The 2024-26 literature genuinely cuts both ways rather than settling: one 2026 study in the International Journal for Educational Integrity found false positives rare across tools, but the same study found Turnitin, GPTZero and Copyleaks failed almost entirely to catch the AI text they were pointed at. A clean false-positive record bought by not detecting anything is not a win.
Universities have voted with their settings. Vanderbilt disabled Turnitin’s AI detector in 2023, reasoning that a 1% false-positive rate across its 75,000 annual submissions meant “around 750 student papers could have been incorrectly labeled,” and still states it “does not offer any institutionally supported AI detection tools.” Pittsburgh cites the “substantial risk of false positives.” UT Austin does not endorse the tools and holds no central contracts. Alabama named “the fallibility of AI-detection tools” as a large part of its decision, and Michigan State, Georgetown, Yale and Boston University have made similar calls. Searching specifically for institutions that reversed course and switched detection back on turned up none. That is a null result rather than proof of a universal, since a quiet re-enabling would never be announced, but the direction of travel is one-way. If your integrity policy treats a detector score as evidence, it is out of step with the institutions that looked hardest at this.
Language learning
10. Duolingo Max
Max is Duolingo’s top tier, and it is alive. It adds Explain My Answer and Video Call, a conversation partner with the character Lily, now joined by a guided beginner version with Falstaff.
We are not quoting a price for it. Duolingo publishes no pricing page (duolingo.com/pricing returns a 404), sells entirely in-app, and geo- and A/B-tests what it shows you. The App Store listing exposes Super price points including $12.99, but no Max in-app purchase is listed at all. Every roundup confidently printing “$29.99 a month” is quoting something it cannot have verified. Open the app and look.
Same on the model: Duolingo’s own Max page still says GPT-4. That copy dates to the 2023 launch and is almost certainly stale, but no newer disclosure exists, so “runs on GPT-4o” is an invention. The real limitation has not changed since launch. Duolingo builds a habit better than anything else on earth and will not make you fluent by itself. duolingo.com
11. ELSA Speak
ELSA scores English pronunciation at the phoneme level and tells you which specific sound you are missing, tuned to the errors typical of your first language. Nothing else in this list does that as well, and the app was updated this month, so it is actively maintained.
Pricing here is unreliable too, and the widely copied “$11.99 a month or $74.99 a year” matches nothing currently on sale. The “Pro” name is retired; the ELSA Pro URL now redirects to a generic subscription page and the App Store lists the tier as ELSA Premium at $19.99 monthly and $99.99 yearly, with several regional variants. ELSA Schools and ELSA Business exist for institutions. No FERPA, COPPA or GDPR claims appear on those pages, which for a tool aimed partly at children is a gap worth asking about.
English only, speech only. elsaspeak.com
Making video and course content
12. Synthesia
Type a script, get a presenter-led video in any of 160-plus languages without a camera. For localising course material across a multilingual cohort, the time saving is real.
Every number in the old version of this post was wrong, so: the entry paid plan is Starter at $29 a month billed monthly ($18 billed yearly), and Creator at $89 monthly ($64 yearly). There is no “Personal” plan. Avatar counts go 9 on Free, 125-plus on Starter, 180-plus on Creator, 240-plus on Enterprise, so the frequently cited “140+” corresponds to no tier that exists.
There is no education pricing. Synthesia’s /education, /solutions/education and /nonprofit URLs all 404, and no student or academic discount appears on the pricing page, which makes it an expensive choice for a school. The avatars also remain slightly uncanny, and students notice. synthesia.io
13. Descript
Descript edits video by editing its transcript. Delete a sentence of text, the video loses that sentence. For a lecturer who is not an editor, it collapses the learning curve to about an hour.
The plan names in circulation are fabricated. There is no “Pro at $30.” The real ladder is Hobbyist $24 a month billed monthly ($16 annually), Creator $35 monthly ($24 annually), and Business $65 monthly ($50 annually). The old “$15 Creator” claim appears to be stale Hobbyist annual pricing.
Descript does something Synthesia does not: it grants discounts to students, educators and nonprofits through an application form. That is the one genuine education angle in this category, and it is worth the ten minutes. descript.com
Accessibility
14. Microsoft Learning Tools
Immersive Reader is free and built into Microsoft 365 apps: text-to-speech, line focus, syllable breaks, picture dictionary, translation. For a dyslexic student it is the single highest-impact free thing in this entire post.
One correction that matters, because it removes a barrier people think exists. Reading Coach is now a standalone free app at coach.microsoft.com, and it needs only a Microsoft account, not a Microsoft 365 Education licence. It runs on the web, Windows and Chromebooks. Every write-up describing it as an M365 Education perk is gatekeeping a free tool.
Do not confuse any of this with Azure AI Immersive Reader, a separate paid developer service. The real problem is discovery: most teachers have no idea these features are already sitting inside the software their school bought. Microsoft Learning Tools
15. Otter.ai
Live lecture transcription that a deaf or hard-of-hearing student can read in real time, and that everyone else can search afterwards.
There is no education plan. That claim has been repeated for years and it is not true. What exists is a 20% discount on Pro for individual students and teachers with a .edu address, taking Pro from $16.99 to $13.59 billed monthly, or $6.67 a month billed annually at $79.99. “Otter for Education” is an institutional offering with no public price behind a sales demo. Business runs $30 monthly or $19.99 per user annually.
Accuracy falls off with lecture-hall echo and technical vocabulary, and no FERPA or COPPA claim was verifiable on their site, which matters if you are recording minors. Alternatives are compared in our AI transcription software roundup. otter.ai
Research tools for students and faculty
16. Consensus
Ask a research question, get an answer synthesised from peer-reviewed papers with each claim linked to its source, plus a meter showing how much the literature agrees.
The plans are Free, Pro and Deep. There is no “Premium” tier at $9.99, and the pricing URL moved, so the old link 404s. Pro is $20 a month billed monthly or $12 a month annually; Deep is $65 monthly or $45 annually. Students and faculty with a school email get up to 40% off, and Consensus says over 170 university libraries partner with it, so check whether yours already covers it before paying.
It only knows indexed literature, so it is weak on very recent preprints and thin outside the sciences. consensus.app
17. Elicit
Elicit builds the literature-review table for you: find the papers, extract methods, sample sizes and findings into columns you can scan. For a systematic review it removes weeks.
Here is the find that pays for reading this section. Elicit’s pricing page defaults to Industry rates, where Pro is $75 a month. Change the dropdown to Academic and a cheaper ladder appears, including a Plus tier that does not exist on the industry side: Plus at $19 monthly or $11 a month billed annually, and Pro at $69 monthly or $39 a month billed annually. Same product, roughly a third of the price, hidden behind a dropdown almost nobody touches.
Extraction accuracy still needs checking against the paper. It is a first pass, not a citation you can trust unread. More options in our AI tools for academic research guide. elicit.com
18. Gemini Notebook (formerly NotebookLM)
It got renamed and almost nobody noticed. NotebookLM is now Gemini Notebook; notebooklm.google is still the address, but the site and its metadata now carry the new name, and the “NotebookLM Plus” brand is gone.
The product is the same and still excellent: upload your sources, and it answers only from them with citations back to the passage. That grounding is why it hallucinates far less than a general chatbot, and why it suits a reading list or a set of lecture notes. Audio Overview turns your documents into a passable two-host podcast, which students with a commute genuinely use.
A free tier survives, now called Gemini Notebook Standard, and it is not a Labs experiment any more. Paid access rides along with Google AI Plus, Pro and Ultra, and qualifying Workspace for Education licences include it. We cannot give you a US price. Google’s plans page is geo-locked and served rupee pricing to every attempt to read it, and forcing US locale parameters did not override it. Rather than copy a figure from an aggregator, check Google’s page from your own country. notebooklm.google
Comparison table
| Tool | Price (billed monthly) | Best for | Standout |
|---|---|---|---|
| Khanmigo | Free for teachers; $4/mo families | K-12 tutoring | Refuses to give answers away |
| Synthesis Tutor | $45/mo | Maths, ages 5-11 | Reasoning over answer-matching |
| Knewton Alta | $49.95/term | College STEM | Routes back to the missed prerequisite |
| MagicSchool AI | Free; $12.99/mo Plus | Teacher admin | 80+ tools, real free tier |
| Curipod | $24/mo | Live participation | Interactive lesson in one prompt |
| Eduaide.AI | $5.99/mo | Budget teacher tooling | Cheapest serious option here |
| Diffit | $14.99/mo | Mixed-ability reading | One text, every reading level |
| Gradescope | Free Basic; institutional quote | Grading at scale | Handwritten maths grading |
| Turnitin | Institutional quote | Similarity checking | Publishes its own caveats |
| Duolingo Max | Not published | Habit-building | Video Call practice |
| ELSA Speak | $19.99/mo Premium | Pronunciation | Phoneme-level feedback |
| Synthesia | $29/mo Starter | Multilingual video | 160+ languages, no camera |
| Descript | $24/mo Hobbyist | Lecture video editing | Education discount exists |
| MS Learning Tools | Free | Dyslexia support | Reading Coach needs no licence |
| Otter.ai | $16.99/mo Pro (20% .edu off) | Live captioning | Real-time lecture transcript |
| Consensus | $20/mo Pro | Evidence lookup | Up to 40% student discount |
| Elicit | $19/mo Academic Plus | Literature reviews | Hidden Academic pricing |
| Gemini Notebook | Free tier; US price unverified | Study from your sources | Answers only from your documents |
What didn’t make the cut
Seven tools were dropped from the previous version. Three are dead:
- Tome. Every path on tome.app returns a Vercel “deployment not found” error. The archive shows Tome had already repositioned as a sales assistant by mid-2024, well before the dates the recycled listicles give, and the last working snapshot is June 2025. The acquisition and successor stories circulating trace only to AI-generated aggregators and are not repeated here. For live options, see AI presentation software.
- Knowji. knowji.com is a parked GoDaddy domain. It was serving real pages as recently as March 2026, and no Knowji apps remain in the US App Store.
- ScribeSense. The domain returns HTTP 200, which is how it keeps appearing in roundups written by scripts that check status codes. The page behind it is a domain-parking stub on Above.com infrastructure. There is no vendor. Why it ended could not be established from any primary source.
Four more were cut for cause rather than death:
- Cognii. The site is up and posted news in January 2026, so it is not abandoned. But it publishes no pricing, its footer copyright stopped at 2025, and its marketing still leans on awards from 2015. Too thin to recommend for a purchase.
- Quillionz. Its “free plan” is a 14-day trial. Its pricing page lists $29.99 and $29.98 for the same quarterly plan and renders a copyright line with the year missing. No shutdown notice exists and it may be perfectly fine, but a vendor that cannot keep two prices consistent on one page is not one to hand a purchase order.
- Calendly. A good scheduling tool that is not an AI education tool. It was padding.
- Poe. Not education-specific, and its own pages contradict each other: poe.com/pricing renders a client-side 404, the plan structure changed to compute-point tiers, and a footnote promising US dollars sits above rupee prices. The only dollar figure Poe itself publishes is “starting at $4.99/month” on its about page.
A pattern worth naming: a dead product does not vanish from search results. It gets recycled by AI-written roundups that never load the page. Status codes lie, parked domains return 200, and pricing gets copied between listicles until it bears no relation to what any vendor charges.
How to choose without wasting a year
Start with the workload that is actually hurting. Grading, differentiation and lesson prep are different problems with different tools, and buying a suite because it covers all three usually means it does none of them well.
Check the free tier’s ceiling before you fall in love. Gradescope’s free plan is real but excludes the AI grading that probably drew you in. MagicSchool’s free tier is real and generous. Quillionz’s “free” is a two-week clock.
Toggle to monthly billing before comparing anything. Most of these pages open on annual pricing, and roughly half the wrong numbers in circulation come from someone reading the annual figure and writing “per month” next to it.
Look for the discount that is not advertised. Elicit hides an Academic column behind a dropdown. Consensus gives students up to 40% off. Descript takes an application. Otter’s .edu discount is real even though its “education plan” is not. Your library may already pay for two of these.
Then the question that outranks all of it: is student data going into this, and who signed for it? If the answer is “the teacher, with a personal card, on consumer terms,” fix that before scaling the pilot. Skills first is a defensible order too, and our roundup of free AI courses costs nothing to work through.
Frequently asked questions
Can teachers use free AI tools with student data?
Usually not safely. Accepting a click-wrap agreement can bind your district to terms nobody reviewed, and the FTC recommends that schools rather than individual teachers decide whether a service’s data practices are acceptable. FERPA’s school official exception also requires the vendor to be under the school’s direct control, which consumer terms permitting model training defeat. The cheapest fix is to put no student personally identifiable information in at all, because no disclosure means no FERPA trigger.
Are AI detectors reliable enough to accuse a student?
No, and the vendor agrees. Turnitin states it “does not make a determination of misconduct” and that a score should start a conversation rather than conclude one. Vanderbilt calculated that a 1% false-positive rate across its submission volume meant roughly 750 papers a year could be wrongly flagged, and disabled the tool. A detector score is a prompt to look closer, never evidence on its own.
Do AI detectors discriminate against non-native English speakers?
The Stanford study found seven detectors misclassified 61.3% of TOEFL essays as AI-written, against about 5% for US eighth-graders, because they were reading vocabulary simplicity as a machine signature. Those were older consumer detectors and Turnitin was not tested. Turnitin’s own research reports no statistically significant bias, but only for documents over 300 words, and it concedes the gap widens below that. Both sides land in the same place: short work by a non-native writer is where this goes wrong.
What is the difference between a teacher seat and a school licence?
A contract. Institutional tiers come with a data processing agreement, admin-controlled retention, SSO and rostering, and a written commitment not to train on your inputs. OpenAI, for example, can execute a DPA for Business, Enterprise and API customers and covers Edu under a separate Student Data Privacy Agreement; Free and Plus get neither. The features look similar. The legal position is not.
Is Khanmigo free for teachers?
Yes. Teacher access is free, and Khan Academy has kept it that way. Families pay $4 a month or $44 a year, and districts get a custom quote. There is no “Khanmigo Plus” tier despite what several roundups claim.
Does COPPA apply to AI tools in the classroom?
For under-13s, yes. The amended Rule’s compliance date passed in April 2026, so it is fully in force. Biometric identifiers such as voiceprints and facial templates now count as personal information, which catches AI proctoring, and vendors must publicly post a retention policy. School consent only covers educational use, never a vendor’s own commercial purposes, and a vendor claiming in its terms that COPPA compliance is the school’s job has it backwards.
Which AI tool is best for lesson planning?
MagicSchool if you want the widest tool set and a free tier that stays free. Eduaide.AI if you are paying yourself, since it does most of the same work for $5.99 a month. Curipod if the problem is a room that will not participate rather than the paperwork.
Is NotebookLM still free?
Yes, though it is called Gemini Notebook now and the free tier is Gemini Notebook Standard. It is no longer a Labs experiment, and paid capacity comes through Google AI plans or a qualifying Workspace for Education licence. Google’s pricing page is geo-locked, so check it from your own country rather than trusting a quoted dollar figure.
The short version
If you teach K-12 and want one thing: Khanmigo for students, MagicSchool for yourself. Both free.
If you are in higher education: Gradescope for grading, Elicit on Academic pricing for reviews, Gemini Notebook for reading. If you support students who struggle to read, Immersive Reader and Reading Coach are free and already installed, and that is the highest-return sentence in this post.
And if you are writing an academic integrity policy, read the Vanderbilt reasoning before you make a detector score mean anything.
Pricing verified against vendor pages in July 2026. Vendors change prices without notice and several geo-target them, so check before you buy. Where a price could not be verified from a primary source, this post says so rather than guessing.