ChatGPT Work vs. Chat and Codex: Scope, Permissions, and Review

When Chat and Work appear in ChatGPT, their difference is more than response length. OpenAI's ChatGPT Work and Codex describes Chat for questions and conversations, Work for multi-step work, and Codex for software development.

Features, plans, usage limits, and surfaces can change. This article therefore does not use a fixed pricing table as its source of truth. Check the current official help and the choices shown in the account doing the work.

Choose by task and review method

Do not expand a one-question task merely because Work is available. A task that combines several source files, a comparison, and a report is a better fit when its intermediate evidence can be reviewed.

Give Work five fields

Deliverable:
Permitted sources:
Prohibited actions:
Validation criteria:
Points requiring my approval:

For example:

Use only the two attached files to create a monthly sales-difference table.
Stop rather than guessing if a column meaning is unclear.
Do not transmit data externally, overwrite the originals, or change integrations.
Reconcile totals to the source rows and show the formulas used.
Save a separate working copy and ask for my review before sharing or sending it.

This format makes a wrong premise visible before it becomes a polished deliverable.

Verify file capabilities in the current surface

OpenAI's document, spreadsheet, and presentation guide explains that availability depends on plan, workspace, file type, and surface.

Do not assume that every account can directly edit every Word, Excel, or PowerPoint file. Test a copy and inspect the output format, formulas, comments, images, and sharing permissions. Do not overwrite the source on the first trial.

Control scheduled and connected work

When Work can run once, on a schedule, after a trigger, or as monitoring, pair the start rule with a stop rule:

• when the task starts;
• which data it may read;
• what happens when nothing changes;
• how many failed attempts cause a stop;
• who approves send, publish, delete, or permission changes;
• which access is removed after completion.

OpenAI's agent guide discusses layered guardrails and human intervention for failure thresholds and high-risk actions.

Review a completed deliverable

AI output can include incorrect facts or fabricated sources. Open the result and verify:

1. source files and date range;
2. recalculated totals, units, and formulas;
3. cited links and whether they support the claim;
4. file differences and removed content;
5. external changes to email, calendars, permissions, or publication state.

A file's existence is not evidence that its contents are correct.

Route the same sales material by its outcome

For “Explain the difference formula,” use Chat and check the calculation yourself. For “Create a difference table and explanation file from two monthly CSVs,” give Work the inputs and output format. For “Integrate this into a monthly program,” give Codex the repository and test criteria.

With illustrative values of 100 last month and 120 this month, the difference is 20 and growth is 20%. Define zero and missing-value handling first: a previous value of zero needs a policy such as “no comparison rate,” not the same division. Check these small expectations in the resulting file. Distinguish permission to create working files from permission to send or publish them.

Conclusion

Use Chat for conversation, Work for multi-step deliverables, and Codex for software development. Give Work the deliverable, permitted sources, prohibited actions, validation, and approval points. Then inspect evidence, calculations, differences, and external changes. The current account interface and official help remain authoritative for plan and format availability.