Cursor for Teachers (September 2026): Pricing, the 18-Year Line, and a Debugging Lesson You Can Actually Run

Tuesday, periods seven and eight, computer lab, the Form 5 programming elective, twenty-four students. On the projector is a script that reads a CSV and calculates a class average, and it will not run. The teacher does not give the answer. He opens the Cursor agent pane and says, as he types, "I'm not asking it to fix this for me. I'm asking it to tell me which line is wrong. Watch it closely — in a minute it's going to get something wrong."

The most valuable thing Cursor does in a school is not write code faster for a teacher. It is put every decision the AI makes on a screen where a class can see it, argue about it, and reject it. This guide covers that, the actual bill, and the 18-year line. The full set of nine tools is in the nine AI tools Hong Kong teachers keep asking about.

50|## The shape of the tool

Cursor is an AI code editor from Anysphere, forked from VS Code. The look, the shortcuts and most extensions carry over, so a colleague with VS Code experience starts at almost zero learning cost. The real difference is that an agent lives inside the editor: it reads the whole project, edits across files, runs commands in the built-in terminal, and then hands every change back as a diff for you to accept or reject one at a time.

That approve-each-diff design is what separates it from terminal-first tools, and it is the one place where it clearly wins in a classroom — students see which line changed, why, and what it looks like when the change is wrong.

The rest worth knowing: models sit in two usage pools, with first-party Grok 4.6, Grok 4.5 and Composer 2.5 in the Cursor Models pool, and third-party models from OpenAI, Anthropic and Google in the Other Models pool at their own API rates. Composer 2.5 is Cursor's own model, released 18 May 2026 and built on Moonshot's Kimi K2.5 open-source checkpoint, tuned for sustained long-running work; it runs only inside Cursor products with no public standalone API, so a school wiring models into its own platform should be looking at OpenAI, Anthropic or a local model instead. Paid plans include unlimited Tab completion. Cloud agents and automations push work off the classroom machine. Bugbot reviews changes agentically, included on Teams and billed by usage on individual plans. Privacy mode is an explicit commitment that neither Cursor nor its providers train on your code. MCP, skills and hooks connect outside tools and let you run your own commands around agent actions.

60|To be direct: **there are no classes, no student rosters, no monitoring view, and nothing connected to curriculum or assessment.** Cursor has never claimed otherwise. Its honest uses in a school are three — programming teachers preparing and demonstrating, teachers building small internal utilities, and working with de-identified student code.

A lesson you can actually run

Back to that double period. The goal is not to fix the script quickly. It is to let a class watch a professional debugging process and see the AI be wrong in public.

A program that breaks on purpose

Before the lesson, put class_average.py in the project with three faults at three different levels:

70|```python

import csv

rows = csv.reader(open("marks.csv")) total = 0 for r in rows: total += int(r[1]) print("Average:", total / len(rows))

The faults: a `csv.reader` object cannot be passed to `len()`; the header row gets counted as a mark; and one blank line in the file crashes the program outright. That is a type problem, a data problem and a defensive-programming problem — and crucially, none of them is a syntax error, so the editor will not underline any of it in red.

    80|Then add a rules file at the project root so the agent knows the lesson's constraints:

```text
This project is a classroom demonstration for a Form 5 programming class.

Rules:
- Python standard library only.
- Raise one problem and propose one minimal change at a time. Never fix several things at once.
- Explain why it is wrong before proposing the change. Explanations in Traditional Chinese (Hong Kong usage); keep code and variable names in English.
- Never rewrite the whole file.

90|"One problem at a time" is the line the whole lesson depends on. Without it the agent will happily repair all three faults in one pass, the class sees a correct answer arrive out of nowhere, and nobody learns anything.

Letting it be wrong in front of the class

Open the agent pane with Ctrl+L (Cmd+L on macOS) and pick Composer 2.5 from the model selector — this work does not need a frontier model. Then reference the file with @:

@class_average.py

Run this file, read me the error from the terminal, then identify only the first problem: which line, and why. Do not change anything yet.

100|Let it run the file itself in the built-in terminal (Ctrl and the backtick key). The class watches the traceback appear, then watches the agent explain that csv.reader returns an iterator rather than a list. That thirty seconds teaches more than ten minutes at the whiteboard.

Now ask for the change. It will probably propose rows = list(csv.reader(...)), the diff pane lights up, and this is the most important moment of the lesson: ask the class "accept or reject?" before you click anything.

On the second round, tackling the header row, there is a good chance the agent proposes next(rows) to skip the first line. Technically correct — and quite possibly without handling the blank line, so the program crashes again a few rows in. That is exactly what you want. Reject a diff on purpose, paste the traceback back into the conversation, ask for "one guard clause only, do not touch other lines", and the class sees something that is very hard to convey by telling: every suggestion the model makes is a proposal, not a ruling.

The conclusion on the board is usually one sentence: this tool is only affordable to someone who can judge a proposal.

What makes Cursor suited to a classroom is not that it writes fast. It is that every change waits for somebody to click accept.

110|## The eighteen-year line

Nothing to do with price, but it decides how a school can deploy it.

Clause 2 of Cursor's terms requires users to be at least 18, or the age of majority in their jurisdiction, whichever is higher. Three consequences follow: a school cannot bulk-create accounts for under-18 learners; passing a teacher's account around the lab is a workaround rather than compliance; and even in Form 6, where some students have turned 18, you end up with one class half compliant and half not, which is no basis for an assessment brief.

On discounts, the official student page says only that anyone can start with Cursor for free and to watch for promotions at on-campus and online events — there is no standing student rate. If a reseller quoting your school claims an official student discount, ask them to point at the Cursor page that says so.

So the division of labour is settled: teachers use Cursor; students use a school platform with gating, monitoring and records. For how that gating works, see student AI safety: four-layer gating, PII protection and teacher monitoring.

120|## Three layers of the bill

Figures below are from the official pricing page and documentation as of September 2026, in US dollars and exclusive of tax.

Plans start at Hobby, free, with limited agent requests, access to Composer and no credit card required. Pro is US$20 a month with both usage pools included and frontier models available. Pro+ is US$60 a month, which Cursor recommends for daily agent users, and Ultra is US$200 a month for power users. Teams standard seats are US$40 per user per month with central billing, SSO, Bugbot and usage analytics; Teams Premium seats are US$120 per user per month with five times the standard seat's agent limits. Enterprise is custom-quoted with pooled usage, SCIM, audit logs and priority support. There is also a Start plan at ₹649 a month, India only, covering the Cursor Models pool alone and not available in Hong Kong.

Model token pricing, per million tokens in and out: Composer 2.5 at US$0.50 / US$2.50; the in-product default Composer 2.5 (Fast) at US$3.00 / US$15.00; Grok 4.6, jointly trained by Cursor and SpaceXAI, at US$2 / US$6; and on the third-party side Claude Sonnet 5 at US$2 / US$10 and GPT-5.6 Terra at US$2 / US$12.

The seat multiplier

130|Putting all 100 staff on standard Teams seats is US$4,000 a month, US$48,000 a year, over HK$370,000. That money buys an editor, not a teaching platform — no classes, no pupils, no curriculum alignment. In practice you buy it for the programming teachers who will genuinely use it, usually two to four people.

Two costs beyond the seat

The first is the Cursor Token Rate. On Teams and Enterprise, selecting a third-party model adds US$0.25 per million tokens on top of that model's own API price, and it applies to included usage, on-demand usage and even bring-your-own-key. Cursor's first-party Grok and Composer models are exempt, which is a clear enough signal that the platform would rather you used its own.

The second is consumption itself. Cursor's documentation is admirably blunt: daily agent users typically total US$60 to US$100 a month, and power users running multiple agents or automations often exceed US$200. Those are per-person figures. A school treating this as a standard tool has to budget by usage, not by multiplying a seat price by headcount.

What it does not do

140|Listing this plainly means nobody arrives at procurement with the wrong expectation.

No native Traditional Chinese interface. As a VS Code fork, a VS Code Chinese language pack translates some inherited menus and panels, but Cursor's own agent chat window, settings pane and account pages remain English. The AI answers in Traditional Chinese perfectly well; the interface is a separate matter, and for colleagues not comfortable in English it is a real barrier.

No education features at all. No classes, no progress tracking, no monitoring, no curriculum alignment. It does not know your panel's rubric or where your scheme of work has reached, so the teacher restates it every session. That is what a school knowledge base is for — see a school knowledge base that makes AI cite your own materials and policies.

A bill that does not follow your financial year. The seat price is a floor and consumption sets the rest, while school budgets are annual and answerable to an incorporated management committee. Per-seat-plus-usage pricing has no brake when a term runs hot.

It is still an editor. For a colleague with no coding background, opening it shows a file tree and a terminal, not a chat box. That barrier is real, and it is not the thing to schedule for a whole-staff training day.

150|## Who to buy it for

The answer is a list of names, not a plan: the two to four programming teachers and IT colleagues who will actually use it, not all 100 staff.

Its home ground is not in dispute. Writing and editing code is first-tier, the VS Code heritage makes onboarding almost free, the approve-each-diff design makes it the best of these tools for classroom demonstration, and first-party Composer 2.5 costs far less than third-party frontier models. If your school teaches programming and someone builds internal utilities, the money is well spent.

The gaps get filled by a different system: gating, monitoring and records on the student side, schemes of work and rubrics in a knowledge base, a fixed number that fits in a budget line, and evidence for the grant. Edor.ai is a per-school annual fee covering roughly 800 students and 100 staff with AI usage included, student conversations pass four-layer gating and teacher monitoring with hint levels set per task, the data belongs to the school and exports in full on demand, and cloud and on-premises deployments are feature-identical. On the grant, the "智" 啟學教 funding asks for evidence of impact, and the PIE module produces it from platform data — see how to use the EDB HK$500,000 AI grant.

Different problems, and they belong in different budget lines.

160|## Common misreadings

• The claim: Cursor is "VS Code with AI", so it can serve as the school's AI strategy. What actually happens: it replaces an editor, not a platform. No classes, no student accounts, no monitoring, no curriculum alignment — and those four things occupy far more of a school day than writing code does.
• The claim: Teams standard seats are US$40 per user per month, so headcount times forty is the annual budget. What actually happens: third-party models on Teams and Enterprise add the US$0.25 per million token Cursor Token Rate, and Cursor's own documentation puts daily agent users at US$60 to US$100 a month in total usage. The seat fee is the floor.
• The claim: if the agent's suggestion passes the tests, accepting it is safe. What actually happens: passing means the program runs. The next() fix in the lesson above is perfectly valid syntax and still crashes on a blank line in real data. Deciding whether to accept remains a person's job.

Three things to try on Monday

170|

1. Install the free Hobby plan and do exactly one thing: open a utility you wrote yourself, ask it to find one problem, and then reject its first diff. Get a feel for the act of approving.
2. Plant one non-syntax fault in your next demonstration program — a type, a data or a boundary problem. It will be the most valuable teaching material in the lesson.
3. Add one line to your panel's written practice: student files shown to an AI carry no names, class numbers or student numbers in the filename or the contents, with the mapping stored elsewhere. A rule short enough to quote is a rule people follow.