ChatGPT Work is the part of ChatGPT designed for doing longer, multi-step work—not just answering a prompt. You give it an outcome, context and constraints; Work can gather information, plan the approach, use approved files and tools, create finished deliverables, and keep a task moving while you review progress and approve important actions.
The important distinction in 2026 is that ChatGPT is no longer one interface for every job. Chat is still the fastest place for questions and short conversational help. Work is for a bounded task with a finished output. Workspace Agents are reusable agents that an organization can build and share for repeatable workflows. Codex remains the dedicated environment for software development and repository work.
Methodology note: this guide is documentation-verified against current OpenAI product pages, ChatGPT Learn, Help Center documentation and release notes, with Work availability, local-file behavior, cloud-browser permissions and credit guidance rechecked on September 25, 2026. It is not presented as a hands-on benchmark. Availability, models, Voice support, credit rates and rollout status can change quickly, so verify the linked official documentation and the controls visible in your account before making a purchase or deployment decision.
Quick answer: when should you use ChatGPT Work?
Use Work when the task has a clear deliverable and needs more than a few conversational turns to finish. Good examples include building a report from several sources, creating or editing a spreadsheet, producing a presentation, researching a topic and turning it into a decision brief, updating a recurring operating review, or carrying a multi-step task across files, plugins and websites.
Stay in normal Chat when you need an answer, explanation, brainstorm, quick draft or small transformation. Switch to Codex when the center of gravity is a codebase, terminal, tests or software-development workflow. Use a Workspace Agent when the workflow should be packaged once and reused by many people, on a schedule or through approved triggers.
| Experience | Best fit | Typical output | Reuse model |
|---|---|---|---|
| Chat | Questions, ideation, short drafts, quick analysis | An answer or conversational result | Start a conversation when needed |
| Work | Longer, multi-step work with a defined outcome | Report, spreadsheet, deck, Site, analysis, completed workflow | One task, project, scheduled task or continuing Work chat |
| Workspace Agents | Repeatable organization workflows shared across a team | Standardized recurring workflow | Build once, share and trigger repeatedly |
| Codex | Software development, repositories, terminals, testing | Code changes, reviews, tests, technical work | Developer-oriented coding sessions and automation |
What ChatGPT Work actually does
OpenAI describes Work as a way to delegate real work to ChatGPT and get a result you can review. Instead of optimizing for a fast reply, Work can gather context, make a plan, execute multiple steps, create files and continue until it has a usable deliverable.
- Admin analytics and outcomes: for workspaces where the Admin Console analytics are available, admins can review active users, credits and token usage across ChatGPT Work and Codex; inspect task-level Insights; track Codex contributions to merged commits and lines of code in Outcomes; and use the Admin plugin or Admin API to analyze trends and build internal reports. This makes Work easier to govern as a business system rather than treating adoption as a simple seat count. OpenAI documented these analytics on September 16, 2026.
- Research and synthesis: gather information from files, connected tools and the web, then turn it into a structured output.
- File creation and editing: create or edit documents, spreadsheets, presentations, reports and analyses.
- Tool use: use plugins and approved connected services for context or actions.
- Browser work: use a cloud browser for supported web tasks or, on desktop, work alongside the built-in browser.
- Longer execution: cloud tasks can continue after you leave the conversation or close your computer.
- Scheduled work: eligible Work tasks can run once, recur on a schedule, use triggers or monitor for changes through Scheduled Tasks.
- Review and approval: you can inspect progress, answer questions, redirect the task and approve consequential actions.
The practical shift is from “help me think through this” to “take these inputs and produce this finished outcome.”
ChatGPT Work vs normal Chat
Chat is still the right default for fast interaction. If the useful result is a paragraph, explanation, list of ideas or a few edits, Work adds unnecessary overhead.
Work becomes useful when the task has a beginning, intermediate steps and a finished artifact. OpenAI’s own guidance gives examples such as briefs, presentations, analysis reports, recurring updates, workflows and files that are ready for review.
| Ask yourself | Use Chat | Use Work |
|---|---|---|
| Do I mainly need an answer? | Yes | Usually no |
| Is there a concrete deliverable? | Optional | Usually yes |
| Does the task need multiple tools or files? | Sometimes | Common |
| Should it continue through several steps? | Usually not | Yes |
| Would a plan and checkpoints help? | Often unnecessary | Often useful |
| Can I review the result at the end? | Conversation is the result | Review is part of the workflow |
ChatGPT Work vs Workspace Agents
These two overlap, but they solve different deployment problems.
Work is an execution experience for a task. You open Work, provide the outcome and context, and supervise the result. It can be a one-off project, a continuing Work chat or a scheduled task.
Workspace Agents are reusable organization assets. OpenAI currently describes them as agents that can be built once, connected to approved tools and company context, shared in a workspace, scheduled, used in Slack or triggered through supported integrations. Admins can set permissions, approval checkpoints and monitoring.
A practical rule: start in Work when you are still figuring out how the job should be done; turn a stable, repeated workflow into a Workspace Agent when multiple people should use the same operating pattern.
ChatGPT Work vs Codex
Codex remains a separate developer-focused environment. OpenAI’s current Help Center says Codex is for writing or debugging code, running tests and commands, reviewing changes and working with repositories. Work is broader: research, analysis, documents, spreadsheets, presentations, reports and other knowledge-work deliverables.
If you previously used Codex for non-programming work, OpenAI now explicitly says you can continue doing so or move that work to ChatGPT Work. Work uses a more general-purpose experience designed around everyday professional tasks, while Codex keeps its repository and developer-tool workflows.
On desktop, Codex also remains a separate view with separate history. Work and Chat live together in ChatGPT’s Recents and can use existing Projects as context.
Local Work vs Cloud Work
This is one of the most important distinctions to understand before handing Work a real task.
| Local Work | Cloud Work | |
|---|---|---|
| Where it runs | Your computer through the ChatGPT desktop app | OpenAI-managed cloud environment |
| Local computer files | Can access approved local files/folders | Cannot directly use files that exist only on your computer |
| Desktop apps | Can use supported local apps with permission | Uses cloud-connected tools, plugins and browser capabilities instead |
| Continue after computer closes | No for work that depends on the local machine | Yes, where the task can run remotely |
| Cross-device continuation | Local execution stays tied to the computer | Cloud Work chats sync across web, mobile and desktop |
| Good fit | Local files, desktop-only apps, work you want to observe on the machine | Research, connected apps, web tasks, background work and scheduled tasks |
OpenAI notes an important privacy nuance: even when Work runs locally, messages and task context may still be stored in the cloud. “Local” describes where the execution happens; it should not be interpreted as a guarantee that the entire conversation stays only on your device.
September 23 update: Voice now works in ChatGPT Work on web and mobile
OpenAI expanded ChatGPT Work on September 23, 2026 by bringing Voice to Work on both web and mobile. You can now speak a task instead of typing it, including requests to create documents, presentations or spreadsheets, use connected apps, or work in a browser. If you end a Voice call while a Work task is still running, the unfinished task can continue in text.
The same release also lets Live Voice use plugins and connected apps on web, iOS and Android. Existing app connections, permissions and usage limits still apply. Free and Go users can use Voice in Chat with the plugins their plan supports, while Voice in Work requires access to both Voice and Work.
The practical change is that Work is no longer mainly a desktop or typed-task experience: your phone can now be a task-launch surface for agentic work. That does not grant extra permissions or bypass workspace controls; Voice uses the same tools and access boundaries already available to the account.
What plans include ChatGPT Work?
Availability is currently in rollout, so there are two layers to the answer.
OpenAI’s current Help Center says Work is gradually rolling out to eligible accounts. It is available on web and mobile for eligible paid plans and in the ChatGPT desktop app when Work is included for the user’s plan and workspace; Business pricing explicitly lists Work across desktop, web and mobile. Cloud and local access can arrive at different times, and workspace permissions can disable specific capabilities.
In practice, the Work toggle and controls visible in your current account are the source of truth for whether a particular surface or execution mode is available to you today.
For organizations, ChatGPT Business explicitly includes ChatGPT Work on desktop, web and mobile. Our ChatGPT pricing guide tracks the broader plan structure, while our ChatGPT Business vs Claude Team comparison covers the team-purchasing decision. For an individual product-level comparison, see ChatGPT vs Claude in 2026.
Plugins are a major part of the Work model
Work becomes more useful when it can retrieve context or take permitted actions in the tools your team already uses. OpenAI’s current plugin system is available across supported ChatGPT and Codex surfaces, subject to rollout and workspace controls. A plugin can package skills, connected apps and app templates; some underlying apps may use MCP or other provider-specific integrations, but those capabilities remain governed by the app and workspace permissions.
Current examples include Google Drive, Gmail, Outlook, Slack, Microsoft 365, Salesforce, GitHub, Canva and many other systems. A plugin may require you to connect an external account, and your plan or workspace administrator can restrict what is available.
This matters because the quality of a Work task often depends more on which trusted context and actions you expose than on writing a longer prompt. A task that needs a current sales pipeline should use the approved CRM source rather than a stale CSV attached three weeks ago.
Cloud Browser: Work can operate on websites
Cloud Browser gives ChatGPT Work its own remote browser session. It can read pages, click, type into forms and carry out supported steps on public or signed-in websites. It does not automatically inherit the accounts signed in on your personal browser; cloud sessions have their own browser state and require separate sign-in when needed.
Cloud Browser is currently available in Work on supported paid plans, excluding Free and Go, subject to rollout and workspace permissions. It can continue after you close your device and pauses when it needs input, sign-in or confirmation.
Website access and consequential actions are separate permissions
Allowing Work to open a website does not mean it can silently complete every action on that site. OpenAI separates site-access permission from confirmation for consequential actions. Work is designed to ask before actions that can create financial, legal, account or other hard-to-reverse commitments, such as confirming a booking or making a payment.
Websites can also block automated agents. A site that works normally in your own browser may refuse access from Cloud Browser, in which case you may need to take over or complete the step manually.
Documents, spreadsheets and presentations
Work is explicitly designed to create editable deliverables rather than only show formatted text in chat. OpenAI documents support for creating and editing documents, spreadsheets, presentations, reports and analyses from instructions, source material or reusable templates.
Where supported, Work can create or edit native Google Docs, Sheets and Slides when the appropriate Google Workspace app is connected and enabled. Microsoft Office is now a separate in-app ChatGPT surface: Word, Excel and PowerPoint use the same ChatGPT Microsoft add-in. In Word, ChatGPT can work with the open document from a sidebar to draft from notes, summarize, revise selected text, and adjust headings, numbering and formatting; PowerPoint is globally available for creating and revising editable decks. These add-ins are distinct from ChatGPT Work itself and follow plan usage limits, workspace permissions and Microsoft 365 admin controls. ChatGPT for Word is available on all ChatGPT plans, including Free, and OpenAI says it will be enabled by default starting October 1, 2026, although admins can still turn it off. For the native Google-side AI experience across Gmail, Docs, Sheets, Meet and Drive, see our Gemini in Google Workspace guide.
The best prompts for file work specify what must stay unchanged: formulas, layout, brand structure, slide order, tone or template rules. That gives the task a reviewable definition of done.
Scheduled Tasks: Work can become recurring
Work does not have to be a one-time session. Eligible tasks can run once, repeat on a schedule, use a supported trigger or monitor for changes through Scheduled Tasks.
Cloud execution matters here: a recurring research or website-checking task can keep running without your computer staying awake. A task that depends on a local folder or desktop app still depends on the local machine and cannot magically turn into cloud execution.
That makes scheduled Work a useful middle ground between a one-off chat and a fully packaged Workspace Agent. If a scheduled task becomes mission-critical or needs to be shared across a team with a consistent configuration, that is a signal to consider whether it should become a more formal agent or automation.
Usage and credits: Work is not a fixed number of messages
Work usage depends on the model, task size, reasoning effort, input size, output length, agents and other work performed. A short local task and a long cloud research job can consume very different amounts of included capacity.
For ChatGPT Business, OpenAI says Standard seats use their included Work/Codex allowance first. If the workspace has purchased credits, eligible usage can continue from the shared credit pool after an included limit is reached. Premium seats provide more included capacity. OpenAI’s current rate card prices additional Work and Codex usage by token consumption rather than a single flat ‘one task = X credits’ rule.
For eligible personal paid plans, credits can also extend supported Work usage after included limits are exhausted. Because model availability and credit rates change, the Usage screen and current rate card are better budgeting sources than an old screenshot or a fixed estimate in a blog post.
A simple way to decide whether a task belongs in Work
Before switching modes, score the task against five questions:
- Is there a concrete deliverable? A report, file, analysis, deck, tracker or completed workflow is a stronger Work candidate than an open-ended question.
- Does the task need multiple steps? Work earns its overhead when it must gather, transform, check and produce—not when one answer is enough.
- Does it need real context? Files, projects, plugins, company sources or websites make Work more valuable.
- Can you define a review point? You should know what a good result looks like and what must be checked before it is used.
- Is this still a one-person task? If many people need the exact same workflow repeatedly, consider graduating it into a Workspace Agent or conventional automation.
Good first Work tasks
- Turn a folder of source material into a structured decision brief.
- Build a weekly operating review from approved project, KPI and team-update sources.
- Create a spreadsheet from raw data, preserving a defined template and formulas.
- Research a market or vendor set and produce a comparison with citations.
- Convert meeting notes, documents and project context into a launch plan.
- Create a presentation from an existing brief and approved brand template.
- Monitor a low-risk public website for changes and summarize what changed.
What not to delegate first
The fact that Work can use browsers, plugins and tools does not make every task a good automation candidate. Start with work where errors are visible and reversible.
- Do not make your first task an unsupervised financial transaction.
- Do not give broad access to sensitive files when a narrow project folder will do.
- Do not let a browser task make legal or binding commitments without review.
- Do not assume a plugin should have write access merely because read access is useful.
- Do not schedule a consequential task and then stop reviewing its runs.
The same principle we use in our 7-Factor Automation Scorecard applies here: choose high-value work that can fail safely, limit permissions, and keep a human review point where the consequences justify one.
A practical first-run checklist
- Pick one bounded deliverable. Avoid ‘run my business’ prompts.
- Choose local or cloud intentionally. Local for computer-resident files/apps; cloud when background execution and cross-device continuation matter.
- Attach only the context the task needs. Use a project, selected files or approved plugins instead of broad access.
- State constraints and non-negotiables. Name the required sources, output format, template rules and actions that need approval.
- Review the plan. Correct the approach before Work spends time executing the wrong task.
- Review the deliverable and action log. Check facts, formulas, citations, file changes and any external actions.
- Only then repeat or schedule it. A workflow should earn more autonomy by succeeding under review.
This checklist is a researched operating pattern, not a claim that AI-XBlog has benchmarked Work across every platform and plan.
FAQ
Is ChatGPT Work the same as a Workspace Agent?
No. Work is the task-execution experience you open when you want a multi-step outcome and finished deliverable. A Workspace Agent is a reusable organization asset that can package a stable workflow, approved tools and company context for repeated use across a team.
Can ChatGPT Work keep running after I close my laptop?
Cloud Work can. Tasks running remotely can continue after you close your computer and pause when they need your input, sign-in or confirmation. Work that depends on local files or desktop apps still requires the relevant local environment.
Can Work use files on my computer?
Yes, through the ChatGPT desktop app when local Work and the relevant permissions are available. Web and mobile Work cannot directly reach files that only exist on your computer.
What does Auto approve mean in ChatGPT Work?
In Cloud Browser settings, Auto approve applies to website-access requests: ChatGPT can review the URL and only pause when the site appears unsafe. It does not remove confirmation for consequential actions. OpenAI says Work still requests confirmation before hard-to-reverse or real-world commitments such as confirming a booking or making a payment.
Does ChatGPT Work replace Codex?
No. Codex remains the dedicated software-development experience for repositories, terminals, code changes and testing. Work is intended for broader professional tasks and finished knowledge-work deliverables.
Does Work use extra credits?
Work first uses the included agentic allowance on your eligible plan. OpenAI currently says that on Plus and Pro, Codex, ChatGPT Work, Excel and PowerPoint can share the same agentic usage allowance when those features are available. After included limits are exhausted, supported usage can continue with purchased credits where flexible pricing is enabled. Actual consumption depends on the model and task rather than a universal per-message charge.
Bottom line
ChatGPT Work is best understood as the execution layer between a normal conversation and a reusable automation. Use Chat when the conversation itself is the product. Use Work when you can define an outcome, provide context and review a finished deliverable. Move repeated, standardized team workflows into Workspace Agents or a conventional automation. Keep software-development work centered in Codex.
The most useful first Work task is not the most ambitious one. Pick something bounded, reviewable and reversible; give it only the files and tools it needs; inspect the plan; and let the workflow earn more autonomy over time.
Primary sources
- OpenAI: How to connect AI usage to business value
- OpenAI: ChatGPT Work product page
- OpenAI Help Center: ChatGPT Work and Codex
- ChatGPT Learn: Get started with Work
- OpenAI Help Center: Using cloud browser in ChatGPT
- OpenAI Help Center: Creating and editing files with Work
- OpenAI Help Center: Business and Enterprise/Edu credit rate card
- OpenAI: Workspace Agents
- OpenAI: Business pricing and feature comparison
- OpenAI Help Center: ChatGPT for Word
- OpenAI Help Center: ChatGPT for PowerPoint
- OpenAI: ChatGPT release notes — Voice in Work and plugins in Voice (September 23, 2026)
Source check: September 24, 2026. ChatGPT Work is actively rolling out; availability, supported surfaces, models, plugins, Voice support, usage and credit pricing should be treated as living information.
AI-XBlog Weekly Brief
Keep up with AI that actually works
Join the AI-XBlog Weekly Brief for major AI updates, practical workflows, useful tools, and editor’s picks. No daily noise.
Reader discussion
Join the discussion
Have you tried this tool or workflow? Share your experience, corrections, or questions. Useful reader feedback may help us improve this article.
All comments are reviewed before publication. Your email address will not be published. Promotional links and low-value spam are removed.
