ChatGPT now offers three overlapping ways to get things done: Chat, Work, and Codex. The simplest rule is also the most useful:
- use Chat when you mainly need an answer, explanation, draft, or brainstorming partner;
- use Work when you need a multi-step process to end in a finished deliverable;
- use Codex when the centre of the task is a codebase, terminal, tests, or software changes.
The mistake is assuming the most agentic option is automatically the best one. Work can research across files and connected apps, create documents, spreadsheets, presentations, reports, and Sites, and continue through several dependent steps. That power is valuable—but it also adds more setup, permissions, review, and usage than a normal chat needs.
Feature check: Product capabilities and availability were checked on July 26, 2026. OpenAI says Work is rolling out to eligible paid accounts, so the exact interface, plan access, and usage limits may differ by account or workspace.
The Practical Difference
OpenAI describes Chat as the place for quick conversational help, Work as an agent for longer multi-step tasks and finished deliverables, and Codex as the dedicated software-development experience.
That distinction is clearer when you focus on the unit of completion:
| What you need at the end | Best starting point | Why |
|---|---|---|
| An answer, explanation, rewrite, idea list, or quick web search | Chat | The result can be produced in one conversational pass |
| A researched report, spreadsheet, slide deck, document, Site, or recurring workflow | Work | The task contains several dependent steps and may need files, apps, tools, or approvals |
| A code change, bug fix, test run, repository review, or technical implementation | Codex | The work must understand and modify software rather than merely discuss it |
There is overlap. Chat can analyse files and write substantial drafts. Work can perform technical research. Codex can explain code. The decision should depend on where the task becomes difficult—not which button sounds most advanced.
Use Chat When the Conversation Is the Work
Chat is usually the better choice when you can describe the request clearly and evaluate the response immediately.
Good examples include:
- explaining an unfamiliar concept;
- rewriting an email or article section;
- comparing two options;
- summarising a document;
- brainstorming titles or plans;
- doing a quick calculation or web search;
- improving an existing prompt.
A useful test is: Could a strong answer in the next message complete most of the task? When the answer is yes, starting an agent often creates ceremony rather than leverage.
Chat also makes correction fast. You can say “make it shorter,” “show the assumptions,” or “use a warmer tone” without waiting for a longer workflow to inspect tools, gather files, and build an artifact.
Use Work When the Deliverable Requires a Process
Work becomes worthwhile when the output depends on multiple connected actions that would otherwise require you to keep moving information between tools.
OpenAI says Work can research and analyse information, work across connected apps and files, and create finished materials such as documents, spreadsheets, presentations, reports, and Sites. It can also run through Scheduled Tasks for recurring work.
Work is a strong fit when at least three of these conditions are true:
- The output is a finished artifact, not just advice.
- Several steps depend on earlier results.
- The task needs multiple files, sources, or connected apps.
- You expect to review progress and redirect the work.
- The same workflow will run again.
- Manual copying between tools is a meaningful part of the job.
For example, “What should a customer-feedback report contain?” is a Chat question. “Read these survey exports, remove duplicates, group the findings, calculate the main patterns, create charts, write the report, and turn it into a presentation” is a Work task.
A Better Way to Brief Work
An agent performs better when the request defines the end state and the boundaries, not merely the topic.
Use a structure like this:
Create a six-page customer-feedback report from the attached files. Separate verified findings from interpretation, show the calculation behind every percentage, cite the source file for each major claim, and include a one-page action plan. Do not contact anyone or publish anything. Stop for approval before creating the final presentation.
This tells the agent what “done” means, how evidence should be handled, and where human approval is required.
Use Codex When Failure Is Measured by Tests and Code
Codex is the better home when the task must operate inside a software project.
Typical Codex work includes:
- tracing a bug through a repository;
- implementing a feature;
- running tests and commands;
- reviewing a pull request;
- updating dependencies;
- changing infrastructure or deployment configuration;
- explaining how a codebase behaves before editing it.
A coding question does not automatically need Codex. “Explain a JavaScript closure” belongs comfortably in Chat. “Find why this React application leaks event listeners, fix it, run the relevant tests, and show the patch” belongs in Codex.
The dividing line is whether success requires changing and validating a real technical environment.
When Work Is Overkill
Work is probably the wrong starting point when:
- the task can be completed in one or two messages;
- you do not yet know what you want;
- the source material is small enough to paste directly into Chat;
- there is no real deliverable beyond a recommendation;
- the task contains sensitive information that does not need to leave its current system;
- you will spend longer reviewing the agent’s actions than doing the task yourself;
- the work is primarily software development and Codex has the more relevant tools.
Agentic work has an invisible coordination cost. You must provide files, permissions, stopping rules, output requirements, and review criteria. That cost is justified when the workflow is genuinely multi-step. It is wasteful when you only needed a clear paragraph.
The Cost Question: Measure Saved Coordination, Not Tokens Used
OpenAI says Work follows the same general usage structure as Codex. Eligible accounts receive included usage, and credits may extend supported features after plan limits are reached. The precise usage varies with task complexity, context size, tools, and output length rather than following one fixed price per task.
That makes “How many credits did it use?” an incomplete measure. A better comparison is:
Value of Work = manual coordination removed − setup, review, and correction time
A long agent run can be economical when it replaces repeated research, file handling, formatting, and artifact creation. A short agent run can be wasteful when a normal chat would have answered the question immediately.
Before delegating, estimate four numbers:
- How long would the full task take manually?
- How much of that time is repetitive coordination?
- How long will it take to prepare the agent’s context?
- How long will a responsible review take?
Use Work when the second number is large and the third and fourth stay controlled.
A Five-Minute Decision Framework
Before choosing a mode, answer these questions:
1. Is the desired output an answer or an artifact?
Choose Chat for an answer. Consider Work for an artifact.
2. Does the task need to act across several sources or tools?
One document usually fits Chat. Several files, apps, and dependent steps favour Work.
3. Is software the thing being changed?
Use Codex when the real output is code plus validation.
4. Can you define a safe stopping point?
Work is most useful when you can specify approvals such as “draft but do not send,” “prepare the issue but do not publish,” or “show the spreadsheet before creating slides.”
5. Would failure be obvious?
Good agent tasks have inspectable outputs: calculations, citations, tests, diffs, tables, or files. Avoid broad delegation when a polished-looking result could hide an incorrect assumption.
Permissions and Review Still Matter
Work can use connected apps and files when those capabilities are available and permitted. That makes least-privilege access important.
Only connect the sources required for the task. Remove unnecessary personal, financial, medical, customer, or confidential information. Explicitly state which external actions require approval. Review citations, calculations, recipients, sharing settings, and final files before anything leaves the workspace.
The point of an agent is not to remove human judgment. It is to move human judgment to the places where it matters most: defining the goal, setting boundaries, checking evidence, and approving consequential actions.
The Bottom Line
Use the smallest tool that can reliably finish the job.
Chat is faster for thinking and conversation. Work is stronger when a multi-step process must become a finished artifact. Codex is purpose-built when the task must understand, change, and validate software.
Work is worth using when it removes real coordination—not merely because it can stay busy for longer.
