AI Guides › Workbench
By Nigel Guy · 7 min read
Most people open ChatGPT, stay in Chat, and paste a forty-page job into the box. Chat gives back a tidy summary and a draft that looks finished. It isn't, because Chat answers and drafts. It doesn't go off, gather your files, build the spreadsheet and come back with something to review. The other two modes do that, and they sit one toggle away.
The rule: pick the mode by what you want back. An answer means Chat. A finished file or a task carried through means Work. A change to code means Codex. Decide before you type, because Work and Codex draw on a separate, metered allowance.
OpenAI's help centre describes the split like this. Chat is for "fast, conversational assistance". Work is "an agent designed for longer, multi-step work and finished deliverables". Codex "remains dedicated to software development and technical work". At time of writing the plans compare like this:
| Mode | What it does | Cost / free tier (at time of writing) | Best for | The catch |
|---|---|---|---|---|
| Chat | Back-and-forth answers, search, brainstorming, short drafts, file summaries | Free plan includes unlimited text chats on GPT-5.6 Luna, within abuse guardrails | Questions, rewrites, advice, deciding what to ask for | It won't carry a multi-step job through to a file you can use. The newest Work models (GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna) aren't offered in Chat |
| Work | Plans a task, uses your files, plugins, browser and tools, then returns a document, deck, spreadsheet, report or Site | Free and Go: limited, desktop app only. Plus and Pro: desktop, web and mobile | Deliverables that would take you real time by hand, and recurring or triggered tasks | Shares a metered allowance with Codex. A big job can use up your five-hour limit quickly |
| Codex | Writes and debugs code, runs tests and commands, reviews changes, works in a repository | Free and Go: limited. Plus: included. Pro: expanded | Software work where you need to see diffs, the shell and Git | Shows developer detail. On desktop its history is kept apart from ChatGPT history |
On prices: OpenAI quotes Plus at US$20 a month and Pro tiers at US$100, US$200 or US$500. I couldn't confirm the sterling price including VAT for UK buyers, so check the £ figure on the checkout page before you subscribe.
Chat feels shallow on big jobs because it isn't built to hold a task open. It won't keep reading sources, building a file, checking that file and revising it. Use Chat to work out what you actually want, then pass that brief to Work.
To use it: on desktop, select ChatGPT from the top-left menu, then Chat from the toggle at the top of the page. On mobile, choose Chat from the dropdown at the top of the screen.
Fill in the task and the materials you have:
You are helping me scope a task before I hand it to an agent. Do not do the task itself.
Task I have in mind: [ONE_SENTENCE_TASK]
Materials I can supply: [FILES_OR_SOURCES]
Who will use the result: [AUDIENCE]
Steps:
1. Ask me up to three questions about anything missing that would change the result. Wait for my answers.
2. Tell me which mode fits: Chat (an answer is enough), Work (a finished file or a multi-step task), or Codex (code changes). Give one reason.
3. If it is Work or Codex, write a brief I can paste in, with these headings: Outcome, Sources to use, Sources to ignore, Constraints, What good looks like, Stop and ask me before.
Before you answer, check that the brief names a concrete deliverable and at least one point where the agent must stop for my approval. If either is missing, fix it.
Work is the agentic mode. It can research, read files, use installed plugins such as Google Drive, Slack or SharePoint, create documents, presentations, spreadsheets and PDFs, and run scheduled or event-triggered tasks. On the web it runs in a cloud environment. In the desktop app you may also be able to run it locally on files you approve.
To use it on desktop, select ChatGPT, then Work from the toggle. On web or mobile, select Work. To point it at a specific plugin, type @ followed by the plugin's name.
Here are the numbers behind the hype. Work and Codex share one usage pool. OpenAI publishes estimated local messages per five-hour window for Plus, and they vary widely by model. GPT-6 Astra allows roughly 5–45, GPT-6.1 Sol 15–160, and GPT-6 Luna 350–3,000. Weekly limits may also apply. Cloud tasks can use more than local ones. Fast mode uses included usage at 2.5 times the standard rate, and image generation uses it about 3–5 times faster. These are OpenAI's estimates, not guarantees. Check your own usage dashboard.
Fill in the deliverable, sources, audience and stopping point:
Role: you are completing a piece of work for me to review, not giving advice.
Deliverable: [FILE_TYPE_AND_LENGTH, e.g. eight-slide deck / one-tab comparison sheet]
Audience: [WHO_READS_IT]
Use only these sources: [FILES_OR_@PLUGINS]
Do not use: [EXCLUDED_SOURCES]
Constraints: [DEADLINE_STYLE_TEMPLATE_WORD_LIMIT]
Order of work:
1. List the sources you can actually open and any you cannot. If a source I named is unreachable, stop and tell me.
2. Draft the deliverable.
3. Mark every figure or claim with its source. Flag anything you inferred rather than found.
4. Stop before sending, sharing, deleting or editing anything outside this chat, and ask me first.
Output: the file, then a short note covering sources used, gaps, and assumptions.
If any input above is blank, ask me for it rather than guessing. Before you finish, open the file and check every slide, tab or section against this brief.
Codex is for repositories, terminals, tests and diffs. In the desktop app, select Codex from the top-left menu. Codex Cloud runs coding tasks on OpenAI's machines, so they keep going while your computer sleeps. You need to create and publish a cloud environment on desktop or web before you can pick it for a mobile task. OpenAI says Work and Codex overlap and that you can keep using Codex for documents if you prefer. The main differences are presentation: Work hides Git and shell details, while Codex shows diffs and review views.
Fill in the repository, the change and how to test it:
You are working in [REPO_OR_FOLDER]. Goal: [CHANGE_IN_ONE_SENTENCE].
Done means: [TEST_COMMAND_OR_OBSERVABLE_BEHAVIOUR] passes.
Do not touch: [FILES_OR_AREAS_OFF_LIMITS].
Steps: read the relevant files and summarise your plan in five lines or fewer; make the smallest change that meets the goal; run [TEST_COMMAND]; show me the diff.
Ask before installing packages, changing configuration, or pushing anything.
If the goal or test is unclear, ask instead of guessing. Before replying, confirm the tests ran and say whether they passed.
Ask yourself three questions, in this order: