education

Best AI Tools for Teachers

The best AI tools for teachers in 2026 compared on price, features and evidence — the free educator workspaces from OpenAI, Anthropic, Google, Microsoft and Khan Academy, the purpose-built platforms including MagicSchool, Brisk, Diffit, SchoolAI and Eduaide, the grading and assessment layer, and an honest account of what AI marking actually gets right.

Updated August 2026

Quick answer: The best AI tool for most teachers in 2026 is ChatGPT for Teachers, which is free for verified US K-12 educators and school staff, gives unlimited access to OpenAI’s flagship model in a workspace built to support FERPA requirements, and costs a school nothing to deploy. Teachers outside the United States, and anyone who wants a purpose-built interface rather than a blank chat box, should start with MagicSchool (free tier, $99.96 per year for Plus) for breadth of teacher tools, or Brisk Teaching if the work happens inside Google Docs, Canvas or Word, because Brisk is a browser extension that reaches into documents you already have rather than a site you have to visit. For adapting a text, article, PDF or video to different reading levels, Diffit does one job better than any general-purpose assistant. Khanmigo for Teachers is free worldwide in English and is the only major teacher assistant funded by a donated infrastructure partnership rather than a subscription business.

The caveat that matters more than any ranking: AI grading is not accurate enough to set a final grade, and not one vendor in this comparison publishes the two numbers that would prove otherwise. The most-cited independent study of LLM essay scoring found roughly 40% exact agreement with trained human raters against roughly 50% agreement between humans themselves (Tate et al., AERA 2024), and an ETS Research Institute reanalysis of the same dataset using GPT-4o found scores averaging 0.9 points below human raters, exact matches only 30% of the time, and significantly lower scores for Asian and Pacific Islander students. Use AI for the first draft of feedback. Keep the grade.

What changed for teachers in 2026

Three things reshaped this market in the last twelve months, and all three cut against the way the category is usually written up.

First, the frontier labs made teacher access free and turned it into a land grab. OpenAI launched ChatGPT for Teachers on 19 November 2025, free for verified US K-12 educators. Anthropic followed with Claude for Teachers on 14 July 2026, also free, also verified US K-12 only. Google made Gemini in Classroom free to every Workspace for Education account at ISTE on 30 June 2025 and has been adding to it since. Khanmigo for Teachers has been free since 20 May 2024 on donated Microsoft infrastructure. A teacher in a US public school can now assemble a full AI stack for $0.

Second, and directly because of that, the independent teacher-tool vendors are retreating from selling to individual teachers. SchoolAI stated outright on 15 July 2026 that it has no self-serve paid plan and cut its free tier to five Space sessions per year. Curipod has removed its individual Premium tier — pricing is now Free or Custom. Wayground, formerly Quizizz, lists only Basic and School-and-District on its US pricing page. Gradescope and Class Companion have no individual paid tier at all. The teacher with a credit card is being squeezed out in favour of the district with a purchase order, which matters if you are the only person in your building who wants these tools.

Third, the evidence caught up, and it is more sober than the marketing. The Walton Family Foundation-Gallup Teaching for Tomorrow study, fielded 18 March to 11 April 2025 with 2,232 US public K-12 teachers, found six in ten teachers using AI, 32% using it at least weekly, and weekly users estimating savings of 5.9 hours per week — about six weeks across a 37.4-week school year. The 5.9-hour figure is the one every vendor quotes. The figure they do not quote comes from the same study: grading and feedback is the task where AI improves quality least, with only 57% of teachers reporting better quality of work on grading and feedback, against 74% reporting better quality on administrative work. And in the strongest randomised evidence yet published, NBER Working Paper 35620 (issued August 2026) ran a two-year cluster randomised trial of Khanmigo across 18 Tennessee middle schools and concluded that the maths gains “resemble those from Khan Academy practice without AI assistance.”

Read those together and the honest 2026 position is this: AI is a genuine time-saver on preparation, materials and admin, an unproven assistant on grading, and not yet a demonstrated driver of student outcomes.

Best AI tools for teachers ranked (August 2026)

Prices read from each vendor’s own pricing page on 19 August 2026, with one exception: Microsoft’s education pages could not be opened directly during research, so the $18 academic figure is Microsoft’s published price as indexed rather than as read on the page, and should be confirmed with a reseller. Where a vendor publishes no figure, this table says so rather than repeating a third-party number — an important distinction in this category, because the AI-generated listicles that dominate these search results carry a large volume of prices that no vendor has charged for years.

The free educator workspaces from the AI labs

These are the highest-value products on the page, because for a large share of teachers they cost nothing and replace several paid tools at once.

ChatGPT for Teachers — free, and the strongest general assistant

ChatGPT for Teachers is a dedicated workspace, not a discount on a consumer plan. It is free for anyone currently employed by an accredited K-12 school or district in the United States — teachers, staff, administrators and district leaders — with eligibility confirmed through the third-party verification service SheerID. It is explicitly not for students.

What you get: unlimited messages with OpenAI’s flagship model, file uploads, web search, image generation, data analysis, memory, deep research, voice, custom GPTs, shared projects for co-planning with colleagues, and connectors that pull in Google Drive, Microsoft 365 and Canva. Anything shared with the workspace is not used to train OpenAI’s models by default. School and district leaders can claim a domain, bring staff into one workspace, and secure it with role-based access controls and SAML single sign-on.

One live inconsistency worth knowing before you plan around it. OpenAI’s product page states in three separate places that the plan is “free for U.S. K-12 teachers and school staff through June 2028”. OpenAI’s own pricing page still says “through June 2027”, and the November 2025 launch post — which also still names GPT-5.1 Auto as the model, superseded well before now — was never updated. Both the June 2028 and the June 2027 end dates were live on OpenAI’s own properties on 19 August 2026, and no separate announcement of the extension appears to exist. The product page is the more current source, but a district planning a multi-year rollout should get the end date in writing.

The launch cohort covers roughly 150,000 teachers and staff across 16 districts including Dallas ISD, Houston ISD, Fairfax County Public Schools, Fulton County Schools and Prince William County Public Schools, plus the Delaware Department of Education statewide.

Limitation: United States only. There is no equivalent free tier for teachers in the UK, Australia, Canada or the EU, who fall back to the standard ChatGPT free plan or a personal subscription.

Related for 2026: OpenAI launched ChatGPT for Teens on 18 August 2026, automatically placing users estimated or stated to be aged 13 to 17 into a guardrailed experience with Study Mode, responsible-homework reminders that redirect assignment-shortcutting into tutoring, and parent-set Study Hours. OpenAI is explicit that this supports learning “outside of the classroom” while ChatGPT for Teachers remains the school-managed path. The practical effect for teachers is that a student’s personal ChatGPT account now defaults toward tutoring behaviour rather than answer-generation, which shifts the homework-integrity calculation somewhat in your favour.

Claude for Teachers — free, and the best writer in the set

Anthropic launched Claude for Teachers on 14 July 2026, giving verified US K-12 educators free access to premium Claude capabilities plus a library of education-specific resources. The supporting legal documents — the US K-12 Terms of Service and the US K-12 Data Processing Agreement — both carry an effective date of 14 July 2026 and are the most substantive privacy commitments any lab has published for this audience.

What the K-12 terms actually commit to, in Anthropic’s own drafting: Anthropic may not train models on customer content; the customer owns the outputs; Anthropic is designated a School Official with legitimate educational interest under FERPA; the DPA rules out selling or sharing student data, commits to 48-hour breach notification, deletion within 30 days of termination, and an annual third-party audit. That FERPA School Official designation is the specific thing district counsel asks for, and it is unusual to see it written into a free consumer-facing offer.

Claude’s strength for teachers is the same as its strength everywhere: long documents. Feeding it a full unit plan, a set of curriculum standards, a district scope-and-sequence and a 40-page article and asking it to build aligned materials is where it outperforms the alternatives.

Two things Anthropic has not published, both flagged here rather than guessed: the exact end date of the free period does not appear on any Anthropic page, and the model powering the offer is not named. Claude Opus 5 shipped on 24 July 2026, ten days after the Teachers launch, so the launch-day model was necessarily an earlier one.

Gemini in Classroom — free on every Education edition, with one important exception

Google’s position is the most generous on paper and the most carefully bounded in practice. On Education Fundamentals, which is free to qualifying institutions, teachers get Gemini for Education with access to Gemini 3, Gemini Notebook (the rebranded NotebookLM), image generation, Guided Learning, interactive quizzes, and Gemini in Classroom across what Google describes as 30-plus use cases. AI rubric generation in Google Classroom completed its rollout on 5 August 2026 and is free on all editions.

The exception in Google’s free tier is the feature closest to actual marking. AI-suggested feedback on written assignments — the “help me write” private comment that drafts feedback on a student’s submission — requires Education Plus or the Teaching and Learning add-on. It was announced on 19 February 2026 and finished rolling out on 30 March 2026. That single gate cuts directly against the “Gemini in Classroom is free” headline, and it is the gate most teachers will hit first.

Google’s published education pricing as of 19 August 2026:

Two 2026 licensing changes matter for budgeting. Google abolished the old “one free staff licence per four paid student licences” model, so every active Fundamentals user must now be covered, with a floor of 50 licences — effective 1 October 2025 for new customers and 1 February 2026 for renewals. And the former Gemini Education Premium add-on at $30 per user per month was withdrawn from sale on 28 August 2025, with Gemini Education renamed Google AI Pro for Education.

Separately, Gemini in Google Classroom expanded to students of all ages — announced 4 August 2026, web rollout from 10 August — where it had previously been restricted to users aged 18 and over. Administrators must now opt out per organisational unit for under-18s rather than opt in, which is a change worth checking if your district has a policy on student AI access.

Microsoft 365 Copilot in Education

For schools already standardised on Microsoft, Copilot Chat is free when signed in with a school account, and the Learning Accelerators — Reading Coach, Reading Progress, Search Coach and Speaker Coach — are free tools inside Microsoft Teams for Education. Microsoft 365 Copilot itself, the version embedded in Word, Outlook, Excel and Teams, is offered as an academic SKU at $18 per user per month for educators, staff and students aged 13 and over, against a $30 commercial list price. Note that Microsoft 365 A1, the free tier for verified educational institutions, includes neither Copilot nor the desktop apps.

Because Microsoft’s own pages could not be opened directly during research for this page, treat the $18 figure as Microsoft’s published academic price as indexed on 19 August 2026 and confirm it with your reseller before budgeting.

Khanmigo for Teachers — free, worldwide, and the only nonprofit in the set

Khanmigo for Teachers is, in Khan Academy’s own words, “now completely, 100% free” — and unlike the OpenAI and Anthropic offers, it is free to educators around the world in English, currently across 44 or more countries, not just the United States. Khan Academy is direct about why: “Thanks to our partnership with Microsoft, we are able to provide free access to Khanmigo for Teachers in English to educators around the world.” Microsoft donated the infrastructure, and Khanmigo for Teachers runs on Azure OpenAI Service. Google announced on 21 January 2026 that Khan Academy is integrating Gemini models starting with Writing Coach, so the underlying model layer is now multi-vendor.

The teacher tier includes roughly 25 activities: standards-aligned lesson planning tied to Khan’s content library, learning objectives, rubrics, exit tickets, quiz and question generation with answer keys, student groupings, lesson hooks, multilingual family emails, class newsletters, report card comments and an IEP Assistant.

Two limitations to know before you build a workflow on it. First, teachers cannot grant students access — classroom access for students is only available through a school or district implementation. Second, Khanmigo does not grade. It builds rubrics and summarises student progress; a human still applies the score.

Paid tiers: learners and parents pay $4 per month or $44 per year (the parent plan covers up to ten children, and a US billing address is required). District pricing starts at $10 per student per year for up to 1,000 licences, self-serve by card; above 1,000 licences the price is not published.

The strongest independent evidence about any tool on this page concerns this one, and it is mixed. NBER Working Paper 35620 by Philip Oreopoulos and Nina Low, issued August 2026, reports a two-year cluster randomised trial across 18 Tennessee middle schools with Khanmigo configured to coach rather than answer. Assignment raised maths achievement by 1.3 national percentile ranks per term, roughly 0.06 to 0.08 standard deviations over a school year, with the implied effect of a full year of active participation reaching 0.14 standard deviations. The authors’ own summary is the line to read twice: “These gains resemble those from Khan Academy practice without AI assistance.” The diagnosis is engagement, not capability — 96% of students tried Khanmigo at least once, but the median student messaged it on only a third of the days they practised, and in only 17% of the exercise sessions in which they made a mistake. Khan Academy’s chief learning officer Kristen DiCerbo has separately acknowledged that only about 15% of students with access regularly engage with it.

The purpose-built teacher platforms

These exist because a blank chat box is a bad interface for a teacher with 25 minutes of planning time. They wrap the same underlying models in named tools — “generate a rubric”, “write an IEP goal”, “make this text easier” — and the good ones save real time on the prompt-writing.

MagicSchool — the broadest toolkit

MagicSchool offers more than 80 teacher tools and more than 50 student tools, covering lesson planning, assessment writing, differentiation, IEP drafting, behaviour intervention ideas, parent communication and, added through 2026, song and podcast generation. Pricing is free with generation caps, $12.99 per month, or $99.96 per year (billed as $8.33 per month), with custom Enterprise pricing for schools and districts. It has repositioned itself in 2026 as “the AI operating system for schools” and was named to the TIME100 education companies list on 29 April 2026. Hillsborough County Public Schools rolled it out to 7,000 educators in April 2026.

Its most creditable 2026 decision is a safety one. On 24 February 2026 MagicSchool retired “Raina”, its AI assistant persona, stating the change was made “to reduce companionship risk”, and added AI disclosure to students and chatbot guardrails. That is an unusually direct vendor acknowledgement of parasocial risk in a classroom product, and it is the right call. Its marketing has not entirely caught up — a testimonial reading “We became best friends with Raina” was still live on the homepage on 19 August 2026.

Watch the integrations claim. MagicSchool’s integrations page lists six items — Canvas, Schoology, Google, Clever, ClassLink and Microsoft — four of them Enterprise-gated, and does not mention its Chrome extension at all. The homepage claim about integrating with Google Docs is not supported by any row on that page. Its Microsoft integration is single sign-on only.

Certification conflict, flagged rather than smoothed: MagicSchool’s trust page describes its ESSA evidence as Level III in prose with a named reviewer, while the badge graphic reused across the homepage and integrations page carries alt text reading Level IV. Use Level III. Privacy posture is otherwise the strongest in the independent set — iKeepSafe FERPA certification and COPPA Safe Harbor, SOC 2 Type II, 1EdTech certification, and a 95% Common Sense Media privacy rating.

Brisk Teaching — the one that comes to you

Brisk is a Chrome and Edge extension first and a web app second, and that architecture is the whole argument for it. It works inside Google Docs, Slides, Forms, Classroom and Drive; Word, PowerPoint and OneDrive; Canvas SpeedGrader, Schoology, Blackboard, Turnitin and plain PDFs. If your marking already happens in SpeedGrader, Brisk is the only tool here that puts feedback generation in the same window.

Its standout capability is Inspect Writing, a video-style replay of how a student’s document was actually written in Google Docs — showing edits, deletions and pasted-in text. Nothing else in this comparison does that, and as a response to the AI-integrity problem it is far more defensible than a detector score, because it shows process evidence rather than a probability.

Brisk announced Curriculum Intelligence on 22 January 2026, keeping generated materials aligned to what a school is actually teaching, and it is now sold as a third district tier. The site claims more than 2,000,000 educators.

Pricing is the problem. Brisk publishes a “Free Forever” tier and two district tiers, and no individual paid price at all. Its Educator Pro page remains live and the plan is still cancellable in-app, which reads like a legacy tier being quietly retired rather than a current product. Privacy posture is strong: SOC 2 Type II with the October 2025 report publicly linked, 1EdTech certification, and a 93% Common Sense privacy rating. Note that Brisk describes that 93% as the highest score among AI tools while MagicSchool describes its 95% as top of the field; both link verifiable pages, and Brisk’s claim does not hold.

Diffit — the best single-purpose tool on this page

Diffit takes a topic, a URL, a PDF, a YouTube video, an image or an excerpt from a novel and produces the same content at multiple reading levels, with summaries, vocabulary, comprehension questions and one-click export to Docs, Slides, Forms, Word or PowerPoint. For a mixed-ability class or an EAL cohort, this is the highest time-saving-per-dollar tool in the comparison.

Pricing: free with 90 days of history, $14.99 per month or $149.99 per year. School and district pricing is tiered by enrolment and quote-only.

Diffit has shipped more consistently through 2026 than anything else here — a dated changelog runs to 11 August 2026 (Front of Room Slides), with a Chrome extension launched 7 August 2026 and an Anthropic connector on 14 July 2026. Its privacy position is the most absolute in the set: “We do not collect ANY student data.” The trade-off is real — Diffit holds no SOC 2 attestation and no 1EdTech certification, so it will not clear a strict district procurement checklist even though it collects less than the tools that will.

SchoolAI — supervised student AI, if a district is paying

SchoolAI is the only tool here built primarily around students using AI under live teacher supervision. Its Mission Control gives real-time visibility into every student’s conversation, with critical-alert routing that is never paywalled, plus a Class Mastery Score and Smart Groups. It earned ESSA Tier 3 certification on 23 April 2026, reviewed by Instructure, from a two-year Jordan School District study of 82 teachers and roughly 14,000 conversations reporting a 28% critical-thinking gain.

Its free tier was cut hard on 15 July 2026 to five Space launches per year, moving Presentation Creator, Flashcards, Mindmap, Video Explorer, extension writing feedback and AI Notetaker behind paid plans. There is no individual paid plan to move up to — the only route is a district quote. Treat SchoolAI as a district purchase, not a personal tool.

One caution on scale claims: SchoolAI’s figures vary by three orders of magnitude across its own materials, from “more than 1 million classrooms” in a press release to “more than 400 school districts” in EdWeek Market Brief to “thousands of schools” on its about page today.

Eduaide.ai — the cheapest serious option

Eduaide has roughly 190 tools at $5.99 per month, the lowest paid price of any full teacher toolkit here. Its Version 3 launched on 3 February 2026, adding evaluators built with Learning Commons, including a Grade Level Appropriateness Evaluator whose rubrics were validated against datasets from Student Achievement Partners, CAST and Achievement Network.

Its privacy answer is the cleanest sentence in this whole category: “Do you collect student data? No we do not collect student data.” It is teacher-facing only. The cost of that simplicity is that Eduaide has no Google Classroom integration, no LMS integration, no Chrome extension and no published compliance certifications — you export a document and take it elsewhere.

Teachmate — the UK and Ireland answer

Teachmate, rebranded from TeachMate AI with the domain moving to teachmate.com, has 162 tools with a centre of gravity that is genuinely different from the US pack: 45 leadership and administration tools and 24 SEND tools, across the England, Wales, Scotland, Northern Ireland and Ireland curricula. If you are a UK head of department or SENCO, this is the tool built for your paperwork rather than an American one you have to translate.

It claims compliance with the Department for Education’s updated Generative AI Product Safety Standards as of 28 January 2026, has partnerships with NASUWT (1 June 2026) and Teach First (3 July 2026), and reports growth to more than 420,000 educators. Its free tier covers 14 of the 162 tools. Paid pricing is quoted in pounds sterling and was not retrievable in a verifiable form on 19 August 2026 — data not available.

The rest, briefly

Almanack has the deepest native Google Classroom integration in the set and a genuinely distinctive feature — report card comments with Voice Mode, backed by a database of official district and provincial report card forms — plus OpenSciEd and Illustrative Mathematics companions. But it contradicts itself on its own site in ways a buyer should notice: its about page claims 150,000-plus teachers while another page claims 500,000-plus; its schools page says “no student data required” while its FERPA page, last updated 11 November 2024, describes collecting basic student information, academic progress data and educational records. Its blog stopped in October 2025.

Curipod does one thing well: write, get instant “Glow and Grow” feedback, discuss, revise, all inside a single period, aligned to 55-plus named curricula. Its free tier is two teaching sessions per week and its individual paid tier no longer exists. Its last blog post was 21 April 2026. It also displays a Common Sense Media logo under a heading reading “Independent review” — what Common Sense actually produced was a Moderate-risk assessment that named Curipod for a fabricated Rosa Parks image, alongside an explicit “privacy: not yet rated”. Curipod’s Common Sense badge therefore implies an endorsement that does not exist.

Twee is the only genuinely language-teaching-specific product here, CEFR-native with roughly 44 tools. Individual plans are $6.50 per month for Basic and $10.50 per month for Pro, both billed annually ($78 and $126 per year respectively), and it is the rare vendor that publishes school pricing openly at $99.95, $89.95 or $74.95 per seat per year by band. It also has no US privacy posture of any kind — no FERPA claim, no COPPA claim, no SOC 2, and a 13-plus age gate.

Nearpod remains a strong interactive-lesson platform, free on Silver (300 MB storage, not the 100 MB repeated across listicles), $159 per year for Gold and $397 per year for Platinum. Its AI Create runs on GPT-4o mini and is English only; the Question Generator is on all tiers, capped at five on Silver, but the Lesson Generator is school and district exclusive. Parent company Renaissance launched Renaissance Intelligence on 19 February 2026, explicitly unifying “previously separate applications” with an AI lesson-creation toolset. No sunset has been announced and nearpod.com is fully live, but that is a consolidation risk worth pricing into a multi-year decision.

Padlet launched Padlet Intelligence in beta on 14 March 2026 (teachers only — students cannot access it) and Piper 2, its game generator, on 13 May 2026. Padlet’s own write-up of Piper 2 is unusually candid, reporting an internal game-quality pass rate rising from 52.5% to 80% while noting that “images are often still incorrect or low quality” and that text-to-speech mispronounces words. Padlet Classroom starts at $160 per year.

Flocabulary has no substantiated AI feature. Lyric Lab is a 2016 creativity tool that is not marketed as AI. It appears on many “AI tools for teachers” lists and should not.

AI grading and feedback: what actually works

This is the section where the marketing and the evidence diverge most sharply, so take the structural facts first.

AI grading is paywalled far more often than the category’s messaging implies. Google’s AI-suggested feedback on written assignments requires Education Plus or the Teaching and Learning add-on. Formative’s Luna feedback requires the $249-per-year Classroom Plan. Wayground’s rubric-based AI grading is district-only. Gradescope’s AI answer grouping is Institutional-licence-only. Kahoot has no AI grading at any price. Across Google Classroom, Formative, Wayground, Gradescope and Kahoot — the five platforms in this section that most teachers already have access to — the marking feature is behind a gate in four and absent in the fifth.

The tools that actually mark work

Gradescope is a different product from everything else here and the distinction matters. Its “AI” is clustering, not generation: it groups identical or near-identical answers into answer groups, you grade one answer per group, and the grade applies in bulk. It reads student handwriting of English text and mathematical notation including fractions and integral signs, with the main constraint being that the answer sits on one line. It handles programming autograding with code-similarity checking and 200-question bubble-sheet templates. Gradescope is the only tool in this comparison that does bubble sheets, code autograding and template-anchored exam OCR. Its Basic plan remains genuinely free and is feature-gated rather than volume-gated, but bubble-sheet and programming assignment types, AI-powered grading, LMS integration and single sign-on are all Institutional-only, and Institutional pricing is quote-only.

Formative’s Luna is the most honest implementation. Grading requires a rubric — Luna can draft one, but a teacher must review and save it first. It grades one student response at a time; Formative’s own FAQ answers “Does Luna automatically grade every student response?” with a flat “No.” It returns suggested rubric levels per criterion with a short explanation for each, and the teacher adjusts, overrides or regrades. It reads handwriting when it arrives as an image inside a free-response answer. And Formative publishes limitations that no competitor matches for candour, including “Fabrication of information: Can generate false information, including made up quotes, citations, and data points” and “It is always important to fact-check responses given by Luna.”

Two things teachers assume are AI graders and are not. Kahoot! does not AI-grade anything: its Type Answer question accepts a maximum of four pre-listed answer strings of 20 characters or fewer with case-insensitive matching, so a correct-but-differently-worded answer scores zero, and open-ended questions are explicitly not scored at all. And Khanmigo generates rubrics and summarises progress but applies no score.

Two of these tools now sit inside larger groups, which matters for a multi-year decision. Formative was acquired by Newsela in June 2023 and its district tier has already been renamed “Balanced Assessment by Formative”; Newsela made three further acquisitions between February 2025 and March 2026. Class Companion was acquired by Panorama Education on 23 April 2025 and is also now sold as Panorama Solara, though classcompanion.com remains live, standalone and free for teachers with no post-acquisition pricing change announced.

What independent research says about AI marking accuracy

This is the material that should decide how you use these tools, and it is consistent across three separate studies covering two independent datasets.

Tate et al. (University of California, Irvine and Arizona State University), presented at AERA 2024. 1,800 middle and high school history and English essays, ChatGPT against trained human raters on a six-point scale with the rubric supplied. Within-one-point agreement was 89% on one 943-essay batch, 83% on 344 English papers and 76% on 493 history essays. Exact agreement was around 40% for AI against around 50% between human raters. Humans were far more likely to award a top grade of 6 or a bottom grade of 1; ChatGPT clustered its grades between 2 and 5. Tate’s own verdict is worth quoting in full: ChatGPT is “roughly speaking, probably as good as an average busy teacher” and “certainly as good as an overburdened below-average teacher”, but “isn’t yet accurate enough to be used on a high-stakes test or on an essay that would affect a final grade in a class.”

Tate also supplies the two questions every buyer should ask a vendor, and which no vendor in this comparison answers: what is the rate of exact agreement between the AI grader and a human rater on each essay, and how often are they within one point of each other?

An ETS Research Institute reanalysis — a signed letter posted to the Hechinger Report’s comments on 10 June 2024 by ETS principal research director Matthew S. Johnson and senior research scientist Mo Zhang, not a formal ETS publication — used the same dataset with GPT-4o across more than 12,000 essays. Scores generated by ChatGPT were on average 0.9 points lower than human ratings and matched human scores exactly only 30% of the time. Essays by Asian and Pacific Islander students received significantly lower scores from the AI than from human raters. The finding that should stop a procurement conversation: “GPT-4o could predict the race/ethnicity of essay writers more accurately than scoring essays. This suggests that the features it uses to predict race/ethnicity may also influence its scoring.”

The Learning Agency’s ASAP 2.0 benchmark, reported by The 74 on 6 May 2025, ran roughly 24,000 human-scored argumentative essays with demographic data attached. Most differences were small; the exception was that Black students received lower scores than Asian students by a margin large enough to warrant attention. But the same disparity appeared in the human scores — the analysis concluded ChatGPT “didn’t introduce new bias, but rather replicated the bias that already existed in the human scoring data.” It independently confirmed the range-compression finding: unlike human graders, who gave out more As and Fs, ChatGPT handed out a lot of Cs, meaning strong writers may not get the recognition they deserve while weaker writing goes unchallenged.

Note the honest conflict between the ETS and Learning Agency findings: ETS finds the disparity running against Asian and Pacific Islander students, the Learning Agency finds it running against Black students relative to Asian students. Different datasets, different models, different scales. Both findings on racial disparity in AI essay scoring should be on the table; neither should be presented as the settled result.

What the vendors publish instead. Only one vendor in this entire comparison publishes an accuracy figure at all: EssayGrader claims “greater than 96% alignment of scores on human graders — based on a study of over 1000 essays”, with no definition of “alignment”, no methodology, no rubric set, no rater design, no link and no publication. CoGrader offers a qualitative claim — grade suggestions are “lower than the usual variance between human graders” — with no metric attached. Gradescope, Wayground, Kahoot, Formative, Google and Khan Academy publish nothing, which is at least honest silence.

Several vendors publish learning-outcome evidence in the space where grading accuracy should be, and it should not be read as the same thing. Wayground’s ESSA Level III Texas study, Kahoot’s 0.72 standard-deviation WiKIT meta-analysis, Formative’s “+6.75 months of additional ELA instruction” and Khan Academy’s “+6.1% next-item correctness” are all outcome measures. None is an agreement-with-human-grader figure.

The practical rule that follows from all of this: use AI to draft feedback, to catch what you missed on the twentieth essay, and to give students a fast formative response between drafts. Set the grade yourself, and say so to students and parents.

AI writing detection: the part to get wrong at your peril

Detection sits inside the teacher workflow whether or not you choose it, so it belongs on this page — and the 2026 evidence has moved decisively.

The decisive study. Van Vlasselaer, Van Droogenbroeck and Spruyt at Vrije Universiteit Brussel published “Who wrote this? Evaluating the reliability of AI detection tools in higher education” in the International Journal for Educational Integrity on 29 June 2026, preregistered and open access. Across 160 documents of at least 4,000 words:

Turnitin classified 100% of the fully AI-generated papers as false negatives. GPTZero, whose educator page headlines “99% Accuracy”, detected none of them either.

The good news in the same study, which deserves equal weight: three of the four tools — Turnitin, Copyleaks and Pangram — classified 100% of the human-written texts as true negatives, with GPTZero the exception, showing a small false-positive rate. The human corpus was written by non-native English-speaking graduate students before 2019, and the authors explicitly contrast their result with the widely cited Liang et al. 2023 Stanford finding that 61.22% of TOEFL essays by non-native speakers were falsely flagged: “Unlike Liang et al. (2023) … we found no false positives for Pangram, Copyleaks, and Turnitin.” On this evidence, current detectors are no longer producing the false positives Liang et al. found in 2023, and are also failing to catch fully AI-generated text. In fairness to Turnitin, its model was updated in October 2025, February 2026 and July 2026, after this study’s test window.

What the vendors themselves concede. Turnitin’s own FAQ, updated 18 August 2026, states its detection “may misidentify human-written, AI-generated, and AI-paraphrased text, so it should not be used as the sole basis for adverse actions against a student.” Its headline claim is a conditional false-positive rate — under 1% for documents with over 20% AI writing — not an accuracy figure, and it publishes no overall accuracy number. Scores between 1% and 19% have been suppressed and displayed as an asterisk since 8 July 2024, because, in Turnitin’s words, “there is a higher incidence of false positives when the percentage is between 0 and 19”. GPTZero has carried a disclaimer since 2023 reading “These results should not be used to punish students”, and its founder Edward Tian told NPR in December 2025: “We definitely don’t believe this is a punishment tool.”

The legal position has hardened. In Matter of Newby v Adelphi University, decided 28 January 2026 in the New York Supreme Court, Nassau County, a student on the autism spectrum was given an integrity violation after his professor cited a Turnitin “100% AI score” on a history essay. Newby said the only help he had received was grammar assistance from a human tutor in Adelphi’s Bridges to Adelphi support programme, and he submitted reports from Grammarly and ZeroGPT that labelled the essay human-written; he also said he was never given a copy of the 100% report. The court held the finding “without valid basis and devoid of reason”, granted the petition, and ordered the record expunged. A second case, Kato v Palo Alto Unified School District, was filed in the Northern District of California in May 2026 over an essay a teacher said Turnitin flagged at 76% AI; the family submitted a 1,162-page Google Docs revision history as evidence the work was written by hand. Palo Alto Unified has denied all the claims, and the case is unresolved.

There has also been consolidation on the detector side. On 23 June 2026, GPTZero was acquired by Superhuman — the company Grammarly renamed itself after acquiring the email client in July 2025 (TechCrunch) — so one of the two best-known AI detectors sold to teachers is now owned by the same company as Grammarly, the writing assistant students routinely cite in their own defence when a detector flags them.

And at least one state has legislated. Oklahoma Senate Bill 1734, the Responsible Technology in Schools Act, was signed on 12 May 2026 after passing the Senate 42-0. It requires a written district AI policy before 2027-28, mandates educator-directed human-in-the-loop use, and prohibits AI from serving as the primary basis for grading, discipline, placement or other high-stakes decisions. The Georgetown FutureEd legislative tracker, updated 13 July 2026, counts 77 bills across 27 states in the 2026 session with 10 enacted.

The Center for Democracy and Technology surveyed 806 teachers of grades 6 to 12 between June and August 2025, publishing in October 2025, and found 43% regularly using AI detection tools, up from 39% in 2024, and 14% reporting a student accused of AI cheating where it could not be proven or was later shown not to have happened. Only 18% of teachers had received any training on detecting AI-generated work.

The defensible practice in 2026 is process evidence, not probability scores. Version history in Google Docs, Brisk’s Inspect Writing replay, in-class writing, oral defence of a submitted piece, and drafts with timestamps all establish authorship in a way no detector output does. If you want the deeper comparison, see our guide to AI detectors.

Pricing compared: what you will actually pay

Three pricing traps specific to this category:

Kahoot’s prices were restructured on 22 October 2025. Every Kahoot price in every listicle — Plus, Premium, Premium+, Pro School and District — is dead legacy pricing that Kahoot itself now describes as supported but no longer updated. The current individual ladder is Bronze $3 per month ($36 per year), Silver $7, Gold $12, Kahoot! One $19 and One Ultra $39, with Silver, Gold and One shown as struck-through promotional prices against $9, $15 and $25 list. Date-stamp any citation.

EssayGrader’s billing period is ambiguous. Its pricing page has a monthly/annual toggle labelled “up to 30% off” but renders only one set of figures, and a testimonial on its own homepage refers to “$19.99/month” — a figure the $14.99 shown sits exactly 25% below, which is what an annual discount looks like. The rendered prices are most likely the annual per-month equivalent; the monthly list prices are not published.

Kahoot’s under-16 age gate catches K-12 teachers regularly. Word cloud, brainstorm and open-ended question types carry the note that, due to data-protection regulations, students under 16 cannot interact with free-form questions without upgrading to Kahoot! EDU. A teacher on a personal Kahoot!+ plan effectively cannot use free-form questions with a Year 9 class.

Best AI tool for each teaching job

Best overall for a US K-12 teacher: ChatGPT for Teachers. Free, unlimited, education-grade privacy, connectors into Drive and Microsoft 365, and shared projects for co-planning. Nothing paid on this page beats free plus the current flagship model.

Best for teachers outside the United States: Khanmigo for Teachers or MagicSchool. Khanmigo is free in English across 44-plus countries; MagicSchool’s free tier is generous and its Plus tier is $99.96 per year. Neither the OpenAI nor the Anthropic free offer is available outside the US.

Best for lesson planning depth: MagicSchool. More than 80 named teacher tools removes the prompt-writing step entirely, which is the actual bottleneck for a teacher with 25 minutes.

Best for working inside documents you already have: Brisk Teaching. The only tool here that operates inside Google Docs, Canvas SpeedGrader, Schoology, Blackboard and Word rather than asking you to visit a website.

Best for differentiation and EAL: Diffit. Any text, any reading level, with vocabulary and comprehension questions, exported in one click.

Best for long-document and curriculum work: Claude for Teachers. Feed it the standards, the scope-and-sequence and the source material together, and it holds all of it.

Best for UK and Ireland: Teachmate. 162 tools built around the England, Wales, Scotland, Northern Ireland and Ireland curricula, with 45 leadership tools and 24 SEND tools that no US product replicates.

Best for exams, bubble sheets and code: Gradescope. No alternative on this page does template-anchored exam OCR, programming autograding or 200-question bubble sheets. The Basic plan is free, but bubble sheets, code autograding and AI answer grouping all require an Institutional licence, which is quote-only.

Best free essay feedback at volume: Class Companion. Free for teachers with no published cap, custom rubrics, handwritten photo upload and student score-dispute resolution. Grade passback to your LMS requires a school plan.

Best paid essay marking for one teacher: CoGrader at $15 per month billed annually for 350 submissions, or Formative at $249 per year if you need per-criterion rubric grading with a per-criterion explanation and a teacher-override record.

Best on a strict budget: Eduaide.ai at $5.99 per month for roughly 190 tools — with the caveat that it integrates with nothing.

Best for schools that need student-facing AI with supervision: SchoolAI, if a district is buying. Its free tier is now too limited for individual use.

Best for classroom visuals and worksheets: Canva for Education, free for verified K-12 educators, and available as a connector inside ChatGPT for Teachers.

Do not buy for AI grading: Kahoot!. Kahoot has no AI grading at any price tier. It remains an excellent review-game tool.

Student data privacy and district procurement

Two questions decide whether a tool survives a procurement review, and they are not the same question.

Does it touch student data at all? The strongest positions here belong to the two tools that collect none: Eduaide (“Do you collect student data? No we do not collect student data”) and Diffit (“We do not collect ANY student data”). Both are teacher-facing only. Neither holds SOC 2 or 1EdTech certification, so neither clears a strict district checklist despite collecting less than the tools that do. That trade-off is real and worth arguing in a procurement meeting.

Does it have the paperwork? On formal certifications, the ranking is clear: MagicSchool (iKeepSafe FERPA certification, COPPA Safe Harbor, SOC 2 Type II, 1EdTech, 95% Common Sense privacy rating) and Brisk (SOC 2 Type II with the October 2025 report publicly linked, 1EdTech, 93% Common Sense) lead the independents. Among the labs, Anthropic’s K-12 Data Processing Agreement is the most explicit, designating Anthropic a School Official with legitimate educational interest under FERPA and committing to 48-hour breach notification and deletion within 30 days of termination. OpenAI states its Teachers workspace is built to protect student data and support FERPA requirements. Google says its Education AI “supports compliance with COPPA, FERPA and HIPAA” — note that is supports compliance, not certification.

Only Class Companion claims Student Privacy Pledge signatory status; no other tool in this comparison mentions it, and the Pledge appears to have lapsed as a meaningful signal.

The one independent assessment of this category is not flattering. Common Sense Media published “AI Teacher Assistants” on 6 August 2025 with an overall risk rating of Moderate, assessing MagicSchool, Curipod, Khanmigo and Gemini for Google Classroom. It found race and background-biased outputs, failures to flag misinformation, and singled out IEP and behaviour-plan generators as among the most concerning features — which is precisely the workflow several of these platforms market hardest. Brisk, SchoolAI, Diffit and Eduaide were not assessed, so their absence from that finding is not a pass.

The practical reading: an AI-drafted IEP goal or behaviour plan is a starting document for a human specialist to rewrite, not an output to file.

How to choose in 2026

1. Check what you already have before you buy anything. If you are a US K-12 teacher, ChatGPT for Teachers and Claude for Teachers are free. If your school is on Google Workspace for Education, Gemini in Classroom is free. If it is on Microsoft 365, Copilot Chat and the Learning Accelerators are free. Most teachers can build a complete stack at $0 and only then work out what is missing.
2. Buy a purpose-built tool for the job the chat box does badly. That is usually differentiation (Diffit), volume marking (CoGrader or Class Companion) or working inside an existing document (Brisk).
3. Prefer the tool that integrates with where the work already lives. A tool you have to visit is a tool you stop visiting in week three. Brisk, Diffit and Almanack are the strongest here; Eduaide, Twee, Curipod and Teachmate integrate with nothing.
4. Do not plan around an individual subscription to a platform that is exiting the individual market. SchoolAI, Curipod and Wayground have all removed or never offered a self-serve teacher tier in the United States, so getting the full product means going through your district. Class Companion also has no paid individual tier, but for the opposite reason: its teacher tier is free with no published cap.
5. Never let an AI set a final grade, and never let a detector score be your evidence. The research is consistent, the legislation is starting to arrive, and the case law is now against it.
6. Ask any vendor selling you AI grading for two numbers: exact agreement with a human rater, and within-one-point agreement. No vendor in this comparison publishes either number. How they respond tells you what you are buying.

Frequently asked questions

What is the best AI tool for teachers in 2026?

For verified US K-12 educators, ChatGPT for Teachers is the best overall option because it is free, gives unlimited access to OpenAI’s flagship model, and runs in a workspace with education-grade privacy controls and an admin console for districts. For teachers who want purpose-built tools rather than a general chat interface, MagicSchool offers the broadest set at more than 80 teacher tools with a free tier and a $99.96-per-year Plus plan. Teachers outside the United States, where the OpenAI and Anthropic free offers are not available, are best served by Khanmigo for Teachers, which is free in English across 44 or more countries.

Is ChatGPT free for teachers?

Yes, for verified K-12 educators and school staff employed by an accredited school or district in the United States. Eligibility is confirmed through the third-party service SheerID, and the plan is explicitly not for students. OpenAI’s product page states the plan is free through June 2028, while OpenAI’s pricing page and its November 2025 launch post both still say June 2027. Both statements were live on OpenAI’s own properties on 19 August 2026, so a district planning a multi-year rollout should confirm the end date directly with OpenAI.

Is Claude free for teachers?

Yes. Anthropic launched Claude for Teachers on 14 July 2026, giving verified US K-12 educators free access to premium Claude capabilities alongside a library of education resources. Anthropic’s US K-12 Terms of Service and Data Processing Agreement, both effective 14 July 2026, commit that Anthropic will not train models on customer content, designate Anthropic a School Official with legitimate educational interest under FERPA, and require deletion of data within 30 days of termination. Anthropic has not published the end date of the free period.

Can AI grade essays accurately?

Not accurately enough to set a final grade. The most-cited independent study, presented by Tate and colleagues at AERA 2024 across 1,800 middle and high school essays, found roughly 40% exact agreement between ChatGPT and trained human raters, against roughly 50% agreement between two human raters. An ETS Research Institute reanalysis of the same dataset using GPT-4o across more than 12,000 essays found AI scores averaging 0.9 points below human ratings, exact matches only 30% of the time, and significantly lower scores for Asian and Pacific Islander students. Tate’s study and a separate Learning Agency benchmark of roughly 24,000 essays both found that AI compresses scores toward the middle of the scale, awarding fewer top and bottom grades than humans do. AI marking is suitable for low-stakes formative feedback and first-pass drafting, not final assessment.

What is the best free AI tool for teachers?

ChatGPT for Teachers if you are a verified US K-12 educator, and Khanmigo for Teachers if you are not, since Khanmigo is free worldwide in English on infrastructure donated by Microsoft. Beyond those, MagicSchool, Brisk Teaching, Class Companion, Gradescope Basic and Canva for Education all have genuinely usable free tiers. Be careful with free tiers that were cut in 2026: SchoolAI now allows five Space launches per year and Curipod allows two teaching sessions per week.

Do AI detectors actually work?

Not reliably. A preregistered study published in the International Journal for Educational Integrity on 29 June 2026 found that Turnitin, GPTZero and Copyleaks each detected 0% of fully AI-generated papers on the study’s strict measure, while Pangram detected 65% on that same strict measure and 97.5% on an inclusive one. The same study found no false positives for Turnitin, Copyleaks or Pangram on human-written text by non-native English speakers, which is a genuine improvement on earlier research. Turnitin’s own documentation states its detection should not be the sole basis for adverse action against a student, and GPTZero has carried a disclaimer since 2023 reading “These results should not be used to punish students”. In January 2026 a New York court set aside an academic integrity finding based on a 100% Turnitin AI score, holding it “without valid basis and devoid of reason”.

How much time does AI actually save teachers?

Teachers who use AI at least weekly estimate saving 5.9 hours per week, which the Walton Family Foundation-Gallup study translates to about six weeks across a 37.4-week school year. That figure comes from a survey of 2,232 US public K-12 teachers fielded between 18 March and 11 April 2025, and it applies only to weekly users — monthly users report substantially smaller savings. The same study found the savings concentrated in preparation, materials creation and administrative work, with 84% of AI-using teachers saying it saves them time on worksheets and activities and 81% saying it saves time on administrative work. On the separate question of work quality rather than time, grading and feedback is where teachers report the least benefit, at 57%.

Can teachers use AI to write IEP goals?

Several platforms including MagicSchool and Khanmigo offer IEP drafting tools, but Common Sense Media’s assessment of AI teacher assistants, published 6 August 2025 with an overall risk rating of Moderate, singled out IEP and behaviour-plan generators as among the most concerning features it evaluated, citing race and background-biased outputs. Treat an AI-generated IEP goal as a starting draft for a qualified specialist to rewrite, never as a document to file. Oklahoma’s Responsible Technology in Schools Act, signed 12 May 2026, also prohibits AI from serving as the primary basis for placement decisions in that state.

Which AI tools work inside Google Classroom?

Brisk Teaching, Diffit, Almanack, SchoolAI, CoGrader, Class Companion, Nearpod, Wayground and Padlet all connect to Google Classroom, and Google’s own Gemini in Classroom is built in and free on every Education edition. Watch for a common gate: grade passback is frequently a paid or district-only feature even when the integration itself is free. Class Companion’s free tier offers Google Classroom single sign-on and rostering but not grade passback, and Wayground restricts grade passback to School and District plans. Eduaide, Twee, Curipod and Teachmate have no Google Classroom integration at all.

MagicSchool vs Brisk: which is better?

They solve different problems. MagicSchool is a destination with more than 80 named teacher tools, best if you want a menu of prompts you can work through, and it publishes a paid individual price of $12.99 per month or $99.96 per year. Brisk is a Chrome and Edge extension that operates inside Google Docs, Slides, Forms, Word, Canvas SpeedGrader, Schoology, Blackboard and PDFs, best if your work already lives in those documents, and its standout Inspect Writing feature replays how a student’s document was actually written. Brisk does not publish an individual paid price. Both hold SOC 2 Type II and 1EdTech certification and both score in the low-to-mid 90s on Common Sense Media’s privacy rating.

Do AI teacher tools comply with FERPA?

Compliance claims vary in strength and should be read carefully. Anthropic’s US K-12 Data Processing Agreement explicitly designates Anthropic a School Official with legitimate educational interest under FERPA, which is the strongest formal commitment among the major labs. OpenAI states its Teachers workspace is built to support FERPA requirements. Google says its education AI “supports compliance with” COPPA, FERPA and HIPAA, which is a weaker statement than certification. Among the independents, MagicSchool holds iKeepSafe FERPA certification and COPPA Safe Harbor status, while Eduaide and Diffit sidestep the question by collecting no student data at all. Several vendors contradict themselves across their own pages on this point, so verify against the current privacy documentation rather than a marketing badge.

Related pages

• Best AI for students — the student-side view, including the free access offers and academic integrity guidance
• Best AI study tools — flashcards, spaced repetition and revision tools your students are already using
• Best AI for homework — what students reach for after your lesson ends
• Best AI detectors — the full comparison, false-positive evidence and what to use instead
• Best AI for presentations — slide generation beyond what Classroom and Canva give you
• Best AI for writing — the general writing tools behind most of these teacher products
• Best AI models — which underlying model powers what, and how they compare