AI Tools for Teachers in 2026: Where Nine Products Actually Sit, Who Their Terms Exclude, and What They Cost a Whole School

Thursday after dismissal, three people left in the staffroom. The IT panel head, handed the school's AI rollout, pushes his phone into the middle of the table with nine freshly installed icons on it: "The principal wants a recommendation by Monday and I still cannot tell you which of these is for planning and which is for writing code." The Form 1 Chinese teacher adds one line: "I only want to know which one I can switch on for my class."

The answers to those two questions shape the whole paper. Of the nine, only two were designed with a class of pupils in mind; the other seven are personal productivity tools for adults.

Three dividing lines

The unhelpful question is which tool is best. Gamma builds layouts, Edpuzzle collects in-video answers and Cursor edits code, and there is no shared metric across the three. What helps is working out where three lines fall.

Who the product thinks you are

Gemini and Perplexity treat you as a consumer: one person, own credit card, accountable to nobody. OpenAI Codex, Claude Code and Cursor treat you as a developer comfortable with a terminal window and version control; all three start at US$20 a month individually, through ChatGPT Plus, Claude Pro and Cursor Pro. Base44 sits between, turning a few sentences of description into a web app with a login and a database. Gamma solves one narrow, genuinely annoying problem: making an outline you have already written presentable.

Only Edpuzzle and Khanmigo were built around a class. One needs to know which pupil answered wrongly at minute seven; the other uses sustained questioning to resist a pupil who just wants the answer. Those two interfaces contain the concept of a class group. The other seven do not.

Terms written for adults

The second line is in the terms of service, and it matters more to a school than any feature difference. Cursor requires users to be 18. Claude also requires 18 and enforces it with an age assurance service rather than a checkbox. The free teacher plans from OpenAI and Anthropic both state that no student accounts are included. Gemini's free student plan is limited to US college students, also 18 and over. Base44 sets its floor at 13, or 16 in the EU. Khanmigo's student side runs through a parent subscription or a district partnership, with paying parents required to be 18 and hold a US billing address, so a teacher cannot enable it alone.

Five of the nine therefore exclude secondary pupils at the level of the contract, and two more require an adult between the pupil and the product. This is what vendors concluded after pricing child-privacy compliance.

The third line is the billing unit: apart from the two class-based tools, everything here is charged per person per month.

How long a free tier lasts

"Has a free tier" and "can finish a real job on the free tier" are different claims, and the four verifiable allowances differ sharply in useful life.

OpenAI Codex offers limited access inside ChatGPT Free, described officially as suitable for very small coding tasks — enough to form an impression, not to finish work. Claude Code is blunter: the Free plan does not include it. Cursor's Hobby plan gives limited agent requests and access to Composer with no credit card, which covers one honest trial at marking a short assignment. Base44 provides 25 message credits and 100 integration credits a month across up to 5 apps, with authentication and a database; a first version is comfortable and three or four rounds of revision exhaust it.

Prices for Gamma, Edpuzzle, Perplexity and Khanmigo could not be verified from an official page here; use each tool's own guide. All figures are US dollars, exclusive of tax.

Matching tool to task

Materials and lessons

To turn teaching logic into a deck, settle the content and sequence with a general assistant first, then hand it to Gamma for layout; asking one step to do both produces something tidy that says nothing. To flip a lesson, Edpuzzle is the only tool here designed for it: questions inside an existing video, confirmation that pupils watched, and the spread of their answers. Perplexity's cited search fits verification during curriculum design, with one lesson to pass on — a citation is not proof the source supports the claim.

Planning, question setting, marking and notices appear on none of these lists, because that work requires knowing your scheme of work, your panel's rubric and your notice format. A general assistant needs all of it restated each time, so the ceiling on time saved is the time spent restating. See a lesson plan in 30 minutes.

Code and small apps

Cursor is a graphical editor and demonstrates best on a projector. Codex is terminal-first and suits automation scripts and pipelines. Claude Code is also terminal-first with steadier Chinese output, which helps when the deliverable is a report for colleagues. All three assume basic programming knowledge, and none is available to pupils.

For a reading log, an event sign-up or a book loan tracker, Base44 is the only genuinely zero-code route. Its billing counts how often the finished app is used, so a tool for five colleagues is cheap while the same tool opened to 200 pupils burns allowance far faster than any estimate.

The multiplication problem

Every price above is per person per month. A school's arithmetic multiplies by 100 staff and then by 12 months, and only then is it a budget line.

A Cursor Teams standard seat is US$40 per user per month, which is US$48,000 a year for 100 people, roughly HK$374,000. A Premium seat at US$120 comes to US$144,000, roughly HK$1,120,000, with third-party models charged separately at US$0.25 per million tokens. A ChatGPT Business standard seat billed annually is US$20 per user per month, US$24,000 a year for 100 people, roughly HK$187,000; a Claude Team standard seat billed annually is also US$20, likewise US$24,000 and roughly HK$187,000. Converted at about HK$7.8 to the US dollar, for estimation only.

Overage sits on top. OpenAI's own documentation puts real Codex usage at an average of US$100 to US$200 per user per month, ChatGPT Business sells credits once included usage runs out, and Claude Team falls back to API rates when the window is exhausted. A per-seat plan does not quote a price; it quotes a floor.

What a school needs is not the cheapest quotation but a number that will not blow up in March.

The gap on the student side

Five of the nine will not take secondary pupils, and that gap does not close because a school ignores it. It reappears with pupils working through a parent's account, a classmate's account, or an account the school cannot see.

Three things have to be built instead. Gating and records: what a pupil asked, whether the answer was handed over, whether anything came up that needs a teacher now, and an interface where a teacher can look and flag — see student AI safety. Knowledge of this particular school: which chapter Primary 5 Chinese has reached, how many bands the composition rubric has, what salutation the notices use, retrieved and cited before an answer is written — see a school knowledge base. And a billing unit that does not scale with headcount.

Records and accountability

The last piece is the workload most schools discover in year two. The EDB "智" 啟學教 grant provides up to HK$500,000 per school, runs to 31 August 2028, and requires interim and final reports on what was done, how many teachers and pupils took part, and what changed.

If whole-school AI use is spread across nine dashboards, five of which have no institutional reporting, that report gets reconstructed from estimates and surveys. See how to use the EDB HK$500,000 AI grant, 12 questions before buying a school AI platform, the model overview for the products' underlying models, and the LLM classroom for the vocabulary.

Common misreadings

• The claim: pick the strongest tool first and sort out the rest later. What actually happens: the strongest options are built for adult developers, so after choosing you still have nothing for pupils. Decide where the student side lives first.
• The claim: roll the free tiers out to all staff and pay only if they run short. What actually happens: free allowances cover a handful of attempts and may be withdrawn at any time. A school process on a free tier rests on a revocable promise.
• The claim: a teacher's personal account is the teacher's own business. What actually happens: student data in a personal account leaves the school with no visibility and no way to investigate, while the responsibility stays the school's.

Three things to try on Monday

1. Open a document with two columns: which AI tools are in use across the school right now, and who pays for each account. It takes one free period and the answer is messier than expected.
2. Take one official pricing page, multiply the monthly per-seat figure by 100 and then by 12, and put that number on page one of the recommendation. It persuades a principal faster than any feature list.
3. Before the next panel heads' meeting, submit one sentence of policy: which categories of student data may never be pasted into an external tool. Every later procurement decision needs that floor.