Skip to lesson
Lesson 07 / 2028 min readUp to date as of 19 October 2026

Codex in practice

Small tasks, usage limits and reviewing results

Prepare a folder with a company brief, task brief and reference photo. Request a draft, check its sources and result. Then explore usage limits and the changes view.

Reading time is approximate. Allow extra time for the exercises; practice estimates are shown with the workshops.

After this lesson
  • Understand what Codex is and read changes (a diff) without jargon
  • Choose a task with a clear scope
  • Check usage limits and reset times
  • Define a review before merging or sending
Section being translated

The chrome and course program are available in English. Deep training modules (prompt library, AI glossary, playbooks) are being translated - some content below is still in Polish.

Edition date and feature availability

Material revised: 19 October 2026; The interface and folder guide was revised on 19 October using supplied screenshots and OpenAI’s public plugin repositories; this does not reverify product access. The models in Chat, Work and Codex were checked on 16 October 2026, and the GPT-5.5 retirement date and Word access rules on 14 October 2026; other prices and availability were not fully reverified. Check the documentation and your account before using a feature: access depends on plan, region and administrator settings. Use the examples for practice and review the results yourself.

See how it works

Small scope → test → next step

  1. Split

    Choose one result rather than the entire project.

  2. Execute

    Provide relevant files and constraints.

  3. Check

    Review changes and run a relevant test.

  4. Continue

    Record decisions and fix errors before expanding scope.

A concrete example

Form fix: reproduce the error → fix → test successful and failed saves → check the user message.

Does a screenshot prove that a form works?

No. Test submission behaviour, including a service failure.

Try it with your materials

A folder is your agent’s desk

Your agent needs context: the information needed for the task. Instead of describing your company every time, prepare one folder with a company brief, a task brief and photos. Point to it again for the next task and check that the agent read the right files.

  1. 01 · Choose the location

    In local Codex, select the project folder. For a cloud task, attach files or provide the repository. Check the visible file list.

  2. 02 · Supply context

    Ask it to read the company brief and task brief, and inspect the specified photo. Request a summary and questions before work begins.

  3. 03 · Review the result

    Save drafts in output/. Compare them with the photo, offer and goal. A finished ad still needs your review.

my-campaign/

Select a file to see its role and contents.

COMPANY_BRAIN.md

Approved facts: what the company does, its offer, audience, websites and voice.

# Company Brain
Updated: [date] | Owner: [person]

## Company
Name: [name]
What we do: [one sentence]
Where we operate: [area]

## Services
[service, scope, source of current pricing]

## Customers
[who we help and what problem we solve]

## Official links
Website: [URL]
Services: [URL]
Contact: [approved contact]

## Brand voice
[three examples of how we write]

## References and evidence
[case studies, approved claims, logo and photo paths]
Photo usage rights: [source and scope]

## Unknowns
[missing information, person responsible]

Never invent prices, results, testimonials or product details.
A folder is not automatic memory. Naming a file COMPANY_BRAIN.md does not load it into every conversation. A local folder is not automatically sent to the cloud. Point to the files in the current task and ask the agent to confirm access. Source materials must not override your instructions or permissions.

Example: an ad from your photo

Add your product or project photo to references/ and record usage rights. In the brief, name the service, audience, format and elements that must stay unchanged. An advertising skill can structure the process; an available image tool can create the visual. Without that tool, the deliverable is a brief and prompt, not a generated image.

Work in the campaign folder I selected or attached. First list the files you can access. Read AGENTS.md, COMPANY_BRAIN.md and brief.md. Inspect the exact reference photo named in the brief; if it is missing or unreadable, ask for it. Summarise the company, audience and goal, and list missing information. Propose a plan before creating materials. Check available tools: do not claim to have generated an image without using an image tool. Save drafts in output/ and list the files created. Do not alter originals or publish anything.

PL/EN templates, without photos or customer data. Unzip, fill the placeholders and add your own files.

Next: build your Company Brain in lesson 15

What Codex is, without the jargon

Codex is an agent that works on a project, most often one with code. It can read files, change them, run commands and tests, and then show exactly what it did. It runs in the ChatGPT desktop app, in a terminal, in a code editor and in the cloud. It shares the same account limits as ChatGPT Work.

If you do not write code you do not have to start with Codex: for documents, spreadsheets and decks, Work is simpler. Codex becomes useful when someone technical is on your team, or when you want to understand how their work looks from the inside.

Six terms that are enough to start

Repozytorium / Repository
A project folder together with the history of every change. Like a document folder where every version is kept.
Diff
A view of differences: what was removed and what was added. Like “track changes” in a word processor, but for project files.
Test
An automatic check that a feature still works. A “green” test means the check passed, not that the program is free of bugs.
Pull request / Scalenie
A bundle of changes submitted for review. Merging approves it and adds it to the main version of the project.
Uprawnienia / Permissions
Settings that say what the agent may do on its own and what it must ask about. Start cautious and widen them once you trust the results.
AGENTS.md
A file with rules for the agent in a given project, e.g. how to run tests and what not to touch. Read at the start of work.

A safe cycle for one task

  1. Goal in one sentence

    “Fix the rounding bug in the price function.” One goal per task. Two goals are two tasks.

  2. Scope

    Name the files or folder and say what must not be changed (e.g. configuration, customer data).

  3. How we will check

    Give the test command or describe the expected result. Without it, “done” only means the agent says so.

  4. Watch and approve requests

    When the agent asks permission for a command or a change outside scope, read exactly what it will do before you click.

  5. Review changes before merging

    Open the diff, read every change and check the test result. The decision to merge always belongs to a person.

Example prompt (fictional project)

The project has a bug: the formatPrice function in [path] rounds a price to one decimal place instead of two.
Goal: fix it and add a test that catches it.
Scope: change only [function file] and [test file]. Do not touch configuration or other files.
Check: run [test command] and show me its output.
Stop after the changes and show me the diff for review. Do not merge or send anything.

Notice that the prompt has a goal, a scope, a way to check and a place to stop. These are the same elements you know from Work, written more precisely.

How to read a diff

A fictional one-line change. A line with − was removed, a line with + was added. The signs matter more than the colours, because colours are not seen the same by everyone.

function formatPrice(cents) {
− return (cents / 100).toFixed(1);
+ return (cents / 100).toFixed(2);
}

We read it like this: instead of one decimal place (toFixed(1)), formatPrice now returns two (toFixed(2)). A price of 12.5 becomes “12.50” rather than “12.5”. If the diff contains changes beyond what the task covered, ask the agent why it made them.

Five questions before you approve

  • Do the changes touch only the files in the task? Extra files need an explanation.
  • Was the test actually run and did it pass, or did the agent only say it would? Ask for the command output.
  • Was anything deleted? Read deletions most carefully.
  • Were new libraries or configuration changes added? Ask why.
  • Are there any passwords, keys or customer data in the changes or the chat? Never paste them into a prompt.
What if I have no code project?

Ask a technical colleague for a small, unimportant practice project, or do the exercise in Chat: paste the fictional function from the diff above and ask for an explanation of what changed and how to check it. The point is not to write code but to build the habit: goal, scope, check, review.

Lesson 07 - ChatGPT Work and Codex in practice

Limits, optimization and the Enterprise tier

In shortCheck account limits and reset times. Start with tasks that have one goal and compare usage after completion.

Below are limits, working practices and questions for your administrator. Update history is in a separate expandable section.

1. What the limits are

Context window

ChatGPT Work doesn't "remember" infinite pages at once (see Lesson 03 on tokens). Split big tasks into smaller packets with a clear goal and named files.

Usage limits

Every plan has an agent-work pool refreshed on a cycle - higher plan = bigger pool. When it runs out, ChatGPT Work tells you plainly and when it returns.

Parallel tasks

How many tasks can run in parallel depends on the plan. In a company, set a queue: urgent, can-wait, and who approves outputs before send.

File access

File scope depends on where the work runs: in the cloud it is uploaded files and connected tools, locally only what you approve on your computer. This is a safeguard, not a limitation. Keep each project in a dedicated folder.

2. Eight rules for optimal work with ChatGPT Work

These are the difference between "fun toy" and "tool that saves hours". Print and pin them by the team's desk.

01

One task = one goal

"Prepare the June sales report" not "handle sales, marketing and by the way the website". Small, specific tasks finish faster and burn less quota.

02

Put context in AGENTS.md once

An AGENTS.md file in the project folder is the standing brief: what the project is, which tools to use, what not to touch. ChatGPT Work reads it on start - you don't repeat it in every prompt.

03

Plan first, execute second

Add: "show me the plan first and wait for approval". You catch a misunderstanding in 30 seconds instead of 20 minutes of agent work.

04

New topic = new session

An old chat drags all its context. Starting new work in a fresh session gives faster replies and slower quota burn.

05

Repeating process = skill

If you're writing the same instruction a third time - turn it into a skill (Lesson 09). Written once, then it runs on a single call.

06

Point at files, don't "look around"

"Data is in /reports/june.xlsx, columns B and F" beats "find the sales data". Every minute of searching costs quota and adds risk.

07

Control the result, not each step

Instead of hovering over the agent, set a checkpoint: draft, diff or a summary to approve. You decide on send - ChatGPT Work decides on the path.

08

Measure what you automate

Log: task, manual time vs. ChatGPT Work time, number of revisions. After a month you know which automations to expand and which to drop.

3. ChatGPT Work Enterprise

The same agent - plus the management layer a larger organization requires.

Team usage rules

Check the limits and available reports for your specific plan with the administrator. A limit shared across products on one account does not automatically mean one pool for the entire team.

Control and audit

The admin decides who accesses ChatGPT Work, which environments and integrations. Audit logs show who ran what - key for GDPR and compliance.

IT-managed environments

Enterprise lets you define which environments the agent runs in, which secrets it sees and which systems it can reach. IT sets it up once - the team uses it safely.

Centrally approved integrations

Instead of 50 people wiring up personal accounts - one approved list of connectors (Slack, GitHub, CRM) with least-privilege access.

What an Enterprise rollout looks like in practice

  1. 1Pilot (2-4 weeks): one team, 2-3 processes, plain Business plan. You gather evidence that it works.
  2. 2Usage policy (1 page): what you may automate, what data may reach the agent, who approves sends. A policy template is in the prompt library.
  3. 3Enterprise conversation with OpenAI: with numbers from the pilot you negotiate terms against real usage - you're not buying blind.
  4. 4Department-by-department rollout: IT configures environments and integrations, each department gets a short training and a list of approved automations.
💡Golden rule: Start with a small task you can check. After repeated trials, measure review time and corrections. The passage of 3 weeks alone does not establish that a process can run without supervision.
More detailChange history and model retirementsDated course information. Open it for an older model or plan change; confirm current terms on your account.
DevDay 2026The recorded DevDay announcement: Pro 200 — half the usage limit, GPT-6 Pro 200 → 100 messages per week, a Pro 500 plan (500 USD/mo). The material describes a transition period and a one-off credit for existing customers. It does not confirm your account's current terms. Check your current plan, limits and actual usage before calculating rollout costs.
Update - July 9, 2026ChatGPT Work + Codex now share one app. OpenAI integrated the work agent into the ChatGPT app - a Chat / Work toggle at the top. Chat = quick conversation. Work = an agent that can grind for hours on your files, apps and projects. Codex remains a specialist mode for developers (code, PR, review). Access: desktop on all plans, web and mobile rolling out on Plus, Pro, Business, Enterprise, Edu.
New GPT-5.6 model family: Sol - flagship, for complex reasoning (code, research, design); Terra - balanced, ~2× cheaper than 5.5, for daily work; Luna - fastest and cheapest, for high volumes. The model picker also shows a reasoning level (Extra High / High / Medium / Low) - higher = smarter and slower, lower = faster and cheaper.
Latest - July 2026ChatGPT Voice in Chat, Work and Codex. Since 23 July 2026 the desktop app (macOS and Windows) lets you talk to the agent: start a task, ask about progress and steer several threads at once. Powered by GPT-Live. On macOS you can turn on "Screen context" to share your active window. Available on Plus, Pro, Business, Edu and Enterprise.
Cheaper and stronger. On 30 July 2026 OpenAI cut prices: Luna by 80%, Terra by 20% - daily agent work got genuinely cheaper. All three models share the same context window (~1.05M tokens, roughly 2,000 pages) and knowledge up to 16 February 2026. There is also a new "ultra" setting - the highest capability mode, splitting a hard task across several agents working in parallel.
New features in Work / Codex. Multi-repository review, image editing inside tasks, an upgraded agent browser and a plugin marketplace (the agent discovers available skills and plugins in your environment).
August 2026What August added. ChatGPT Voice now works in Work and Codex in the desktop app - you start a task by speaking and steer it as it runs. On macOS there are "Appshots": a hotkey attaches an app window to a thread, together with a screenshot and the available text. Plugin discovery improved too - rankings favour plugins people keep using after installing them. (Archive: the GPT-5.6 Sol price promotion until 21 Nov 2026 is superseded - GPT-6 Sol costs half as much.)
September 2026Astra (GPT-6) - a new generation. Astra (GPT-6) supports demanding tasks. Choose a model by testing quality, time and usage on your own process. Check access and limits in your account; API billing and limits are separate. See lesson 13 for a worked cost example.
Latest - 22 Sep 2026GPT-6 Sol and GPT-6 Luna. Cheaper GPT-6 models trained similarly to Astra. Sol (gpt-6-sol, USD 2/10 per million input/output tokens) for everyday work, code and agents - it makes about half the factual mistakes of GPT-5.6 Sol. Luna (gpt-6-luna, USD 0.10/0.50) for fast, high-volume tasks. API prices ~50% below GPT-5.6. Available in Work and Codex on Plus, Pro, Business, Enterprise and Edu; Free and Go get Luna in the desktop app. Start with GPT-6 Sol and switch to Astra when Sol isn't enough.
⚠ Upcoming retirements. From 26 Oct 2026 (planned) you can't create new custom GPTs, and on 11 Dec 2026 existing GPTs stop running - migrate them to plugins (lesson 09). The GPT-5.5 retirement date for ChatGPT, Work and Codex is 14 October 2026; the API is unaffected. Choose an available model: GPT-6.1 Sol or, for Free/Go in the desktop app, GPT-6 Luna.
Archive: August retirements. On 30 Aug 2026 the standalone "DALL·E GPT" is retired - generate images in ChatGPT Images. On 31 Aug 2026 GPT-5.4 and 5.4 mini leave Codex for ChatGPT sign-ins; the replacements are gpt-5.6-terra and gpt-5.6-luna. If your team has 5.4 set as a default anywhere, change it now, before it stops working.

ChatGPT interface guide

Usage: read your account limits

Module 07

In short ↳The screenshot shows two usage windows and renewal times. The 100% and 82% values describe this account when captured.

Where to find it: Settings → Usage

Example interface. Menu labels, available tools and account settings may differ in your version. The numbers refer to the explanations below.

Select the screenshot to enlarge it.

What you see on screen

  1. 1

    Plan limits

    The screenshot describes a limit shared across Codex, Work, agents and Excel, excluding Chat. This does not establish a team-wide pool or the absence of other limits.

  2. 2

    5-hour limit

    Read the percentage and renewal time. Five hours names the usage window, not guaranteed continuous working time. Per the pricing page (18 Oct 2026) Pro plans currently have no five-hour limit - you will mostly see this counter on Plus and Business.

  3. 3

    Weekly limit

    You cannot derive a task count from 82%; tasks differ in duration, tools and settings.

  4. 4

    Usage resets

    No resets are available here. Do not plan work assuming one will become available.

  5. 5

    Credits

    The credits and top-up section begins below. Inspect current terms and cost settings before enabling it.

How to use it - in plain words

  1. Before a large task, inspect both limits and renewal times.
  2. Split work into stages and inspect usage after the first.
  3. When usage is low, save progress and inspect available options.

Try it on your account

Find both renewal times. Explain why 82% does not mean two more large tasks.

More detailDevDay 2026 announcementsRecorded product changes. Check current availability and account terms before making a decision.

DevDay - San Francisco, 29 September 2026

OpenAI DevDay 2026 announcement summary

In shortA record of nine announcements from 29 September 2026: agents, models, plugins and plans. The descriptions reflect that material; check current documentation and your account for availability.
An agent can use tools to carry out a task. For example, it can find a document, extract data and prepare a report. Its access and permissions determine which actions it can take. Review the result before you use it.
Foundation

Agents API - public beta

Not chat over an API, but task execution: hosted runtime, memory, tools and multi-agent work in a single task. Plus Computer Use in the API, Tool Search and automatic Context Compaction for long conversations.

New

Decisions API - a model that decides

Built on GPT-6 Luna. It doesn't generate text, it picks one of the options you define and returns it with a confidence score: classification, routing, the agent's next step. It answers in about 150 ms. Limited preview, pricing has not been published yet - check the current price before you build on it.

Competition

Dots - agents that keep working

OpenAI answers Meta's Muse with Dots: personal agents on GPT-6 Astra running 24/7 on their own cloud computer, connected to more than 4,000 apps. They keep context between conversations and work with the laptop closed. Background app access is read-only - changing data or sending anything needs your approval. One primary Dot for now; Pro and Business Premium (Enterprise/Edu: beta, off by default).

Codex

Codex - security and many environments

Codex Security Cloud scans your repo for vulnerabilities, investigates the findings itself, removes duplicates and prepares the fixes - and scanning can run continuously. Plus multiple cloud dev environments, Codex Cloud working with the laptop closed, voice control in the CLI, the /agents view and code review in the desktop app.

Pricing

Ultrafast - fast, at a price

Codex up to 8x faster (up to 300 tok/s), the API up to 6x faster. OpenAI quotes 6x the standard price - for GPT-6 Astra that works out to 60 USD/M input and 300 USD/M output, though some of those line items are calculated rather than published in the DevDay materials. Available in the API and in ChatGPT Work and Codex for Enterprise and the Pro 500 plan (500 USD/mo).

Plans

Pro 200 and Pro 500 - recorded changes

The DevDay material describes a Work and Codex limit change from 20x to 10x, GPT-6 Pro messages from 200 to 100 per week, and a Pro 500 plan (500 USD/mo) with Work and @ChatGPT for Slack and Teams. It also describes a transition period and a one-off credit for existing customers. Confirm current terms on your account.

New

ChatGPT Space and Pages

A shared space where a team and ChatGPT build pages: interactive documents with charts, checklists and images that update themselves from connected sources. Pro, Business and Enterprise, desktop and web (mobile: view only). Collaborative slides with export to PowerPoint and Google Slides are due within weeks. Team Tasks and the Slack/Teams connection are Business and Enterprise only.

Plugins

Plugin extensions and @ChatGPT in Slack and Teams

Plugins get a home in the ChatGPT sidebar: panels beside the conversation, fullscreen apps and file viewers (web support for Free and Go is still to come). You can mention @ChatGPT in Slack and Teams channels, threads and direct messages; teammates do not need their own ChatGPT licence.

Beta

Meetings plugin - notes without a bot

Takes notes on calls and in-person conversations without a bot joining, creates a summary with action items and adds it to a Space. Beta in the Mac desktop app for Pro and Business. Agree recording consent and GDPR rules with participants before you use it.

Check before you renew: Five useful changes landed in the same release wave: GPT-6.1 Sol is in ChatGPT Work and Codex, the /agents view shows every agent at once, code review moved into the desktop app, the CLI takes voice dictation, and projectless sessions can start outside a project. The GPT-5.5 retirement date for ChatGPT, Work and Codex is 14 October 2026 (API unaffected) - if any of your saved prompts name it, they will need editing.
Models: GPT-6 Astra and GPT-6.1 Sol are available now. GPT-6.1 Sol (API id gpt-6.1-sol) is near Astra on agentic coding, computer use and professional work at 2 USD/M input, 10 USD/M output and 0.10 USD/M cached input - a fifth of Astra's price. It is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; in regular Chat it is expected later.

Check your understanding

An agent changed code. What do you do before accepting it?
A Codex agent says the fix works. What do you check first?
In a diff you see a line with a minus and a line with a plus. What does it mean?

This is a self-assessment. Mark the lesson complete after doing the exercise and reviewing the result.

Do this now

Request one small change on a copy of your site or sample file. Record the original state, check the result and how to revert. No project? Request a local demo page without publishing it and test one button.

How to know you can do it

Keep as evidence: A folder and one reviewed result.

  • You confirm which files the agent read.
  • The company brief, task brief and photo have distinct roles.
  • You check the draft against sources and the goal.
  • For code, review the diff and test result; for an ad, review copy and reference fidelity.

Mark the lesson complete after the exercise and review. Reading progress and quiz answers alone do not confirm practical skill.

Lesson status
In progress — 0% of the page scrolled
Next