ChatGPT Work is OpenAI’s attempt to turn a conversation into a sustained project. Instead of answering one prompt and stopping, Work can break a goal into steps, gather information across connected systems, create finished artifacts, continue for hours, and pause when it needs guidance or approval.
OpenAI introduced the product on July 9, 2026, alongside GPT-5.6 and a redesigned desktop app that brings Chat, Work, and Codex together. The distinction between these surfaces is useful: Chat remains the conversational mode, Work focuses on research and finished deliverables, and Codex focuses on technical implementation.
Work is not simply a longer chat window. Its product model includes connected apps, files, browser and desktop interaction, scheduled execution, progress updates, and artifacts such as spreadsheets, presentations, documents, reports, Sites, and lightweight apps. That broader reach also creates more consequential security and governance questions.
From answers to finished work
A normal assistant can draft a campaign brief. A work agent can gather customer research, locate relevant files, create the brief, derive assets for several markets, and keep context across the sequence. OpenAI describes ChatGPT Work as able to stay with complex projects for hours by decomposing them and completing steps independently.
The user is not expected to disappear from the process. OpenAI emphasizes that people can follow progress, answer questions, redirect the task, and approve important actions. That interaction model is critical: a long-running agent should expose its plan, current state, blockers, and pending decisions rather than behave like an opaque batch process.
Strong use cases share several properties:
- the task spans multiple sources or tools;
- output needs to be a usable artifact, not a paragraph;
- context must carry across several steps;
- progress can be reviewed at checkpoints;
- important actions have clear approval owners.
A one-line factual question does not need Work. A recurring operating review assembled from several systems might.
Connected apps are context and authority
ChatGPT Work can use plugins to connect with systems such as document stores, messaging platforms, email, calendars, project trackers, and CRMs. A prompt can mention a specific app, and ChatGPT may also select relevant connected capabilities based on the request.
Connections do two different things: they expose information and they may expose actions. Reading a project plan is not equivalent to updating it. Summarizing a message is not equivalent to sending one.
OpenAI’s plugin controls preserve the permissions of the underlying app and source system. If a user cannot access a file in the source, a plugin should not grant access through ChatGPT. Workspace admins can configure roles, read and write actions, approval requirements, sync boundaries, and other app-specific settings.
Before enabling a workflow, document:
- which sources the agent may read;
- which actions it may propose;
- which actions it may perform after confirmation;
- what data may appear in generated artifacts;
- how activity is logged and reviewed;
- who owns the final output.
The convenience of one agent crossing systems should not erase their separate security boundaries.
Finished artifacts, not just generated text
OpenAI positions Work as a way to create slides, sheets, docs, analyses, reports, and Sites. This changes the quality bar. A plausible paragraph can be useful during brainstorming; a spreadsheet used in a meeting needs correct formulas, source traceability, and stable formatting.
Every artifact type needs its own validator:
| Artifact | Useful checks |
|---|---|
| Document | citations, headings, claims, links, template compliance |
| Spreadsheet | formulas, units, ranges, totals, missing data |
| Presentation | narrative flow, source labels, legibility, brand rules |
| Site | responsive layout, accessibility, permissions, working interactions |
| Report | methodology, freshness, contradictions, approval status |
ChatGPT Work can refine outputs in the background, but completion should mean “meets the acceptance criteria,” not “the agent stopped editing.”
Scheduled Tasks change the operating model
Scheduled Tasks can run once, repeat on a schedule, trigger when an event occurs, or monitor for changes. OpenAI gives examples such as reviewing Slack updates each week, checking dashboards each morning, monitoring customer feedback, and updating a presentation after new email arrives.
This moves ChatGPT Work into the territory of operational automation. Recurring tasks require controls that ad hoc chats often omit:
- an owner and an expiry date;
- explicit sources and destination;
- a record of the last successful run;
- duplicate suppression;
- failure notification;
- rate and spending limits;
- review when permissions or schemas change.
A scheduled agent should not silently fail for weeks or keep producing an obsolete report after its owner changes teams. Treat the schedule as a production job, even if creating it felt conversational.
Desktop, web, and mobile are not identical
OpenAI launched Work with staged availability. Its announcement said web and mobile rollout began with Pro, Enterprise, and Edu, followed by Plus and Business, while release notes describe paid-plan rollout excluding Free and Go on those surfaces. The updated desktop app was made available globally for macOS and Windows, bringing Chat, Work, and Codex together, with product capabilities and usage still dependent on plan and workspace policy.
On desktop, Work can use local files and desktop apps with permission. A built-in browser supports web research and compatible tools. Mobile can start work, review progress, and continue projects, while the desktop experience can perform richer local actions.
Because availability changes, teams should verify the current plan documentation rather than copy a launch-day entitlement table into permanent internal guidance.
What ChatGPT Work does not solve
Longer autonomy does not eliminate hallucination, stale sources, permission mistakes, or ambiguous ownership. Work may produce a polished artifact whose underlying assumption is wrong. It may also encounter prompt injection in connected documents or websites.
Do not give the agent a broad instruction such as “keep all our systems updated” without defining source priority, actions, and stop conditions. Safer goals are bounded: “prepare a draft status report from these approved sources, flag conflicts, and request approval before sending.”
High-impact decisions still need qualified humans. The agent can assemble evidence and draft a recommendation; it should not become the legal, financial, security, or personnel decision-maker merely because it can work for hours.
A practical rollout plan
Start with a repeatable task employees already understand. Keep the first version read-only. Define success and run it alongside the existing process. Compare completeness, accuracy, elapsed time, and review effort.
Then add one controlled action, with explicit confirmation. Monitor false positives, unexpected tool choices, and access problems. Only schedule the workflow after the interactive version is reliable, and keep an easy way to pause it.
This staged approach reveals whether the agent truly improves the process or merely moves effort from creation to review.
From long-running work to a production media workflow
ChatGPT Work can be the planning and artifact layer around media operations without becoming the media engine itself. Suppose a marketing team needs to turn an approved DOCX brief and several PDF appendices into a clean campaign package. Work can locate the files, summarize differences, draft an index, and prepare the requested output plan.
When a conversion or merge is needed, Claude or Codex can call a specialized external service such as Medux through an authorized MCP workflow. That is a separate production step; it does not mean Medux is natively integrated into ChatGPT Work.
The Claude MCP DOCX-to-PDF tutorial shows the operational pattern: inspect the source, request an upload URL, upload the document, create a conversion task, poll its status, and download the completed PDF. The PDF merge-and-split tutorial applies the same task lifecycle to multiple documents.
A responsible Work project would:
- identify only approved source files;
- propose conversion and ordering without uploading yet;
- ask for approval before external transfer;
- call the media or document tool once with an idempotency key;
- persist the task ID and monitor it at a bounded interval;
- verify file type, page count, size, and naming;
- return a draft package for human review;
- require separate approval before sharing or publishing.
Scheduled Tasks could monitor an approved folder for new briefs, but the default action should be to prepare a review queue, not automatically publish every output. ChatGPT Work is most valuable when it carries context and coordinates the project while explicit tools perform transformations and people retain authority over consequential actions.