OpenAI Codex for Teachers: The Complete Guide (September 2026) — Plans, Credit Pricing and Building a Classroom Utility

Buried in OpenAI's own documentation is a sentence few schools ever see: Codex costs roughly US$100 to US$200 per user per month on average. That is not the ChatGPT subscription — that is the usage on top of it. An ICT panel head who budgets from the seat price will find the numbers do not add up by mid-year.

This guide sets out what Codex actually is in September 2026, the parts a teacher can genuinely use, a complete walkthrough, and the credit rate card a school needs to be able to read. The full set of nine tools is in the nine AI tools Hong Kong teachers keep asking about.

What this tool actually is

First, clear up a common confusion: today's Codex has nothing to do with the 2021 model of the same name. Codex is now OpenAI's agentic coding product line — not a single model but a set of surfaces: a terminal CLI, extensions for VS Code and other IDEs, a web version, a desktop app, an iOS app, automatic code review on GitHub, and Slack and Linear integrations. Underneath it runs the GPT-5.6 family (Sol, Terra, Luna) and GPT-6 Astra, released in September 2026.

It is built for developers. The interface is a command line, the unit of work is a code repository, and the design assumes a user who reads diffs, uses git and knows when to stop the agent. In a school the people who can genuinely use it are ICT teachers, the IT coordinator who looks after internal systems, and teachers automating repetitive administrative data work. It is not a lesson-planning tool and it is not a student platform — OpenAI's own teacher plan says in as many words that it is not for students.

Key features

• Local agent: run codex in a project directory and it reads files, edits them and runs commands on your machine, asking for approval at each step if you want.
• Cloud tasks: hand work to a cloud environment that keeps running after you close the laptop, then pull the result back locally.
• AGENTS.md project instructions: a Markdown file in the repository root setting out rules and conventions, which Codex reads at the start of every session.
• Local code review: /review produces a prioritised list of findings against uncommitted changes, a commit or a base branch without touching your working files.
• Permissions and sandboxing: /permissions decides what Codex may do without asking, and it is the first thing a teacher should configure.
• Scripting and CI: codex exec runs non-interactively, which suits repeatable batch workflows.
• Skills and plugins: package repeated instructions as skills and connect GitHub, Gmail, Google Drive and other tools through plugins.
• Model tiers: Sol for the hardest reasoning, Terra as the everyday workhorse, Luna for fast, high-volume work, switchable at any time with /model.

Full tutorial: building a class grouping and timer tool with the Codex CLI

The scenario: you teach ICT in a secondary school and want a single-file web tool for the school intranet — paste a class list, generate random groups, and run a lesson timer. No server, no internet connection, no student data collected.

Step 1 — check your plan. Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise. Free only covers very small tasks; for real use, start at Plus (US$20 a month).

Step 2 — install the CLI. On macOS or Linux, run the official installer in a terminal (Windows users can use the npm or Homebrew route documented on the official CLI page):

curl -fsSL https://chatgpt.com/codex/install.sh | sh
mkdir class-toolkit && cd class-toolkit
codex

The first run of codex prompts you to sign in; choose "Sign in with ChatGPT" to use your subscription.

Step 3 — set permissions before you start. Type /permissions and pick a mode that requires confirmation before editing files or running commands. This is not a formality: an agentic tool's default instinct is to act, and on a school machine you want to see what it intends to do first.

Step 4 — create the project instructions. Type /init and Codex generates an AGENTS.md. Add the following, and it will honour these rules in every session:

This project builds single-file web utilities for Hong Kong secondary school teachers.

Rules:
- All output must be one index.html file with no external CDN and no network calls.
- Interface text in Traditional Chinese (Hong Kong usage), for example 分組, 計時, 重設.
- No analytics, no cookies, and no code that sends data out of the browser.
- The class list lives only in browser memory and is cleared on refresh.
- Comment the code in Chinese so another teacher can maintain it later.

The last three rules matter most. School-built utilities rarely fail on functionality; they fail by quietly shipping a class list somewhere it should never go.

Step 5 — give it the task. Pick a model (/model and GPT-5.6 Terra is plenty for this) and describe the goal:

Build index.html, an offline classroom tool with two tabs.

Tab one, "Random groups":
- A textarea, one student name or number per line
- Set either students per group or number of groups, whichever the teacher chooses
- A "Group" button that shows each group as a card, plus a "Regroup" button
- An "Export text" button that copies the result to the clipboard

Tab two, "Lesson timer":
- A countdown in minutes with start, pause and reset
- Background turns orange at one minute remaining, red and flashing three times at zero
- Type large enough to read from the back row on a projector

Use HTML, CSS and vanilla JavaScript in one file. When done, tell me how to open and test it in a browser.

Step 6 — review before accepting. Codex shows the file it proposes to create. Read it. Look specifically for fetch, XMLHttpRequest or an external <script src="http..."> — any of those breaks the rules in AGENTS.md, so tell it to remove them. Then run the built-in review:

/review

Choose "Review uncommitted changes" and it returns a prioritised list of findings without modifying your files.

Step 7 — test and hand over. Open index.html in a browser, then unplug the network and try again to confirm it works offline. Once satisfied, pass the file to your IT coordinator for the intranet. For the next tool in the series you can run non-interactively:

codex exec "Following the rules in AGENTS.md, build seating.html: a computer-room seating planner where student cards drag onto desk slots, printable."

Step 8 — watch your usage. /status shows the current model, permissions and remaining allowance. Check it every couple of weeks; that is the only way to learn where your real consumption sits.

Pros and cons

Pros

• One subscription covers the terminal, IDE, web, desktop and iOS rather than charging separately per surface.
• Clear model tiering: Luna's credit rate is one twentieth of Sol's, so routine work costs very little.
• AGENTS.md lets a panel write its conventions into a file that new teachers inherit automatically.
• /review gives a read-only review mode, useful for checking student code or a colleague's script.
• An API key can drive Codex in CI or shared environments, which suits IT automation.
• The most mature ecosystem, with the most Chinese and English learning material, so training costs least.

Cons

• A command-line interface is a high barrier for non-technical teachers and demonstrates poorly on a projector compared with a graphical editor.
• The pricing structure is genuinely complicated: seats, included usage, credits and fast-mode multipliers stack on top of one another.
• The free US K-12 teacher plan is unavailable in Hong Kong, so local schools pay full price.
• No Traditional Chinese interface.
• No class, student or curriculum functionality of any kind.
• Models churn quickly — GPT-5.4 and GPT-5.4 mini retired in Codex on 31 August 2026 — so any workflow a school builds needs maintenance.

Pricing and fee breakdown

Codex has no separate subscription; it comes with a ChatGPT plan. Figures are from the official pricing pages as of September 2026, in US dollars, exclusive of tax.

Credit rate card (credits per million tokens)

OpenAI also states that a typical Codex task on GPT-5.6 Sol consumes 5 to 30 credits, and that fast mode draws credits faster — for GPT-6 Astra, fast mode is charged at 2.5x the standard rate.

What a school should read into this:

The seat multiplier. Putting 100 staff on Business standard seats at the annual rate is US$2,000 a month, US$24,000 a year, around HK$187,000. But that is the wrong way to buy it: of 100 teachers, perhaps two or three genuinely need Codex.

Overage beyond included usage. Once included usage is exhausted, Plus and Pro users buy extra credits and Business or Enterprise workspaces draw on a shared credit pool. OpenAI's own figure of US$100 to US$200 per user per month tells you the US$20 seat is an entry fee, not the bill.

Annual versus monthly. Business is US$20 per user per month annually against US$25 monthly, a 25% difference. Before committing annually, be sure this is a long-term tool and not a one-year experiment.

Student accounts. OpenAI's terms set a minimum age of 13 with parental permission required under 18. But ChatGPT for Teachers states explicitly that it is not for students, and ChatGPT Edu is a university offering quoted by the sales team with no published price. For a Hong Kong primary or secondary school wanting a monitored AI environment for pupils, this route simply does not exist.

Education offers. ChatGPT for Teachers is free through June 2028, but only for serving staff at accredited US K-12 schools or districts; Hong Kong teachers do not qualify. OpenAI also advertises up to a 75% nonprofit discount on ChatGPT Business or Enterprise — whether a subsidised Hong Kong school qualifies is a question for OpenAI.

Compared with the alternatives

Side by side with Edor.ai

To be clear: Codex is much stronger than Edor.ai at writing code and automating work, and we do not compete with it there. What follows compares the other questions a school asks during procurement.

Why a school still needs a school-based platform

Codex knows exactly what it is, and it is explicit about the four things it does not do.

The student side is blank. OpenAI draws the line clearly: the teacher plan is not for students. That is a responsible position, but it means a school wanting pupils to use AI under gating and supervision must look elsewhere. The Edor.ai tutor gives hints rather than answers by default, the hint ceiling is set per task by the teacher, and every conversation is recorded.

It does not know your school. Your scheme of work, your panel's rubric, your notice format, your SEN accommodation process — all of it lives in the school's own documents, not in the model's memory. A school knowledge base makes the AI retrieve those documents and cite them, as described in a school knowledge base that makes AI cite your own materials and policies.

Its cost structure does not fit a school budget cycle. Seats, included usage, credits, fast-mode multipliers: that is a billing model designed for engineering teams, not for an annual budget answerable to an incorporated management committee. We chose a fixed whole-school annual fee, absorbing model price movements ourselves, for exactly that reason.

It will not evidence your outcomes. The "智" 啟學教 funding requires interim and final reports. The Edor.ai PIE module produces the supporting data straight from platform usage — see how to use the EDB HK$500,000 AI grant.

Edor.ai is built specifically for Hong Kong schools: a Traditional Chinese (Hong Kong) interface, Cantonese oral practice, alignment with local curriculum and grant requirements, and per-school rather than per-head pricing. We will not claim to beat Codex at writing code. We do not, and we do not need to. The two address different layers of the problem.

Takeaway

Codex is an excellent developer tool and a worthwhile investment for ICT teachers and IT coordinators — provided you buy it for the two or three people who will use it rather than 100 accounts. And keep OpenAI's own sentence in mind: an average of US$100 to US$200 per user per month in usage, written in their documentation, not ours.

For how the OpenAI model line-up compares on price and capability, see the OpenAI model page; to get the vocabulary straight first, start with lesson one of the LLM classroom.