Use Chat when you want to think something through. Use Work when you want a finished deliverable. Use Codex when you want to change a technical system.
That is the short answer. It is also a useful management distinction: conversation, delegation, and execution are different kinds of work, even when the same AI sits behind them.
OpenAI now places Chat, Work, and Codex together in the ChatGPT desktop app. The interface is new, but choosing well does not require technical knowledge. Start with the result you need.
The Decision Table
| Choose | The question to ask | Good examples | What you should expect back |
|---|---|---|---|
| Chat | Do I need an answer or a thinking partner? | Explain a concept, brainstorm, pressure-test a decision, rewrite a message, summarize a file | A conversational answer or short draft |
| Work | Do I need a completed piece of work I can review? | Build a deck, compare vendors in a spreadsheet, synthesize research, refresh a recurring report | A reviewable document, presentation, spreadsheet, analysis, Site, or workflow |
| Codex | Do I need changes inside code, files, repositories, or a technical environment? | Fix a bug, update a website, run tests, review a pull request, build an internal tool | Inspected and tested changes to a technical artifact |
This matches OpenAI's own guidance: Chat for working through something, Work for carrying a larger task to a reviewable result, and Codex for software or technical tasks.
Use Chat When the Work Is Still Conversational
Chat is the right starting point when you want an explanation, a challenge, a few options, or a short piece of writing. The value is in the exchange. You ask, inspect the response, add context, and change direction.
Good Chat tasks include:
- Explain this proposal at the level of a COO who knows the market but not the technical architecture.
- Pressure-test my reasoning before tomorrow's pricing discussion.
- Give me three ways to structure this message, then ask which direction I prefer.
- Summarize these notes and tell me what seems unresolved.
Chat is also where an unclear assignment can become a clear one. If you know you need help but cannot yet name the output, use the conversation to define the goal before moving into Work.
Do not use Chat as a substitute for delegation when the task requires several sources, many steps, and a polished file. You may get a good outline. You will still be doing the project management by hand.
Use Work When the Result Should Be Ready to Review
ChatGPT Work is designed for a clear outcome that takes meaningful effort to produce. It can gather context from files and plugins, move through several steps, use approved tools, and create reviewable documents, presentations, spreadsheets, Sites, analyses, and recurring workflows.
OpenAI says Work can continue for hours while you follow progress, answer questions, redirect it, and approve important actions. Longer does not automatically mean better. The advantage is that the task can keep its sources, plan, execution, and artifact together.
Good Work tasks share four traits:
- There is a named deliverable.
- The relevant sources are available.
- The quality bar and constraints can be explained.
- A person can review the result before it is used.
Examples include:
- Turn research notes and a campaign template into a launch brief, then adapt the approved version for three markets.
- Compare vendors using a defined scorecard and return a spreadsheet with risks, missing information, and a recommendation.
- Build an eight-slide leadership presentation from approved source documents and flag unsupported claims.
- Review weekly updates from connected project sources, refresh the operating agenda, and stop before sharing it.
OpenAI's getting-started guide recommends telling Work the outcome, sources, constraints, definition of good, and points where it should stop for review or approval.
Do not use Work for every question merely because it is more capable. A short explanation does not need a multi-step task or extra credit use.
Use Codex When the Work Lives in a Technical System
Codex is for software development and technical tasks. It can inspect a project, work with code and files, run commands and tests, review pull requests, and implement changes in local, isolated worktree, or cloud environments depending on the setup.
Good Codex tasks include:
- Find and fix the cause of a failing checkout flow, then run the relevant tests.
- Add a new page to a website while following the repository's design and engineering instructions.
- Review a pull request for correctness, security, and regressions.
- Update a reporting script and verify the generated output.
- Trace a bug across two repositories and propose a contained fix.
Codex can write prose and create non-code files. That does not make it the best default for every business document. Use it when the task benefits from access to a project, filesystem, terminal, version history, tests, or technical tools.
The July 9 desktop release keeps Codex as a dedicated experience inside ChatGPT and adds inline editing, pull request review in the side panel, multi-repository projects, and faster Computer Use powered by GPT-5.6.
The Same Project May Use All Three
Choosing a mode is not a lifetime commitment. A substantial project can move through the surfaces as the kind of work changes. Imagine a leadership team preparing for a board meeting.
Start in Chat to sharpen the thinking
Ask Chat to challenge the narrative, identify likely board questions, surface assumptions, and help define what the meeting needs to accomplish. The result is clarity, not a final deck.
Move to Work to build the deliverable
Give Work the approved narrative, operating data, prior deck, and slide template. Ask for a draft with evidence attached to each major claim and a pause before the final design pass. The result is a reviewable presentation.
Use Codex only for technical changes
If the preparation reveals that an executive dashboard needs a new metric, use Codex in the relevant repository to inspect the data source, implement the change, and run tests. The result is a change to the system that produces the information.
The mistake is asking one surface to carry all three kinds of responsibility without changing the brief or permissions.
A Simple Routing Test
Before you begin, finish this sentence: At the end of this, I need ______.
- If the blank is an answer, perspective, explanation, or short draft, choose Chat.
- If it is a finished file, analysis, plan, or repeatable workflow, choose Work.
- If it is a tested change to software, files, data, or a technical environment, choose Codex.
If you cannot fill in the blank, start in Chat and clarify the assignment.
Prompt Each Mode Differently
A useful Chat prompt
I am deciding whether to move our leadership meeting from weekly to biweekly. Ask me for the context that changes the decision, then help me map the tradeoffs. Do not recommend an answer until we have identified the risks.
The prompt invites a conversation.
A useful Work prompt
Review the attached meeting notes, decision log, and project updates. Create a two-page operating brief for Friday's leadership meeting. Separate decisions, blockers, owners, and unresolved questions. Use only the attached sources, flag contradictions, and return a draft for my review before treating it as final.
The prompt defines a deliverable, sources, rules, and review point.
A useful Codex prompt
Inspect the reporting project and find where the weekly active-user metric is calculated. Explain the current logic, add the approved exclusion rule with the smallest safe change, update or add tests, and run the relevant checks. Do not change unrelated files.
The prompt defines a technical scope and verification standard.
Permissions Should Follow the Task
The three surfaces can appear together while still using different tools, data, execution environments, and controls. A connected Work task does not automatically inherit every Codex permission, and a local Codex task does not make every company data source available.
For business use, begin with the narrowest access that can complete the assignment. Reading and drafting carry less risk than writing to a system, sharing externally, running code, or scheduling future actions. Add explicit review before consequential steps.
The person assigning the work remains responsible for the result. Open the generated file. Check the numbers and sources. Inspect every slide or spreadsheet tab. Ask what the system could not access and which assumptions it made.
The Practical Takeaway
The new ChatGPT app makes AI more capable, but the useful skill is still judgment.
Choose Chat for the conversation. Choose Work for the deliverable. Choose Codex for the system. Give each one the context it needs, the boundaries it should respect, and a definition of done you can actually review.
Start with one real task, not a tour of features. The mode that returns useful work will become obvious quickly.