ChatGPT has spent years waiting for us to type a question. ChatGPT Work is changing that relationship. You give it an outcome, access to the right material, plus enough freedom to carry out the job. It can then research, organize files, build documents, create presentations, update spreadsheets, and work across connected apps.
Although it sounds impressive, it also creates plenty of room for chaos. OpenAI launched ChatGPT Work on July 9, 2026, powered by GPT-5.6 with Codex technology built in. It is available to Plus users across desktop, web, and mobile, though the desktop app goes further, as it can work with local files, apps, websites, and browser-based tools.
OpenAI describes Work as an agent that can remain with a project for hours, breaking a large job into smaller parts before producing the finished material. The best results still depend on what you give it. A vague request creates vague Work, only now it can waste 40 minutes instead of 40 seconds.
Use Work for jobs, not questions.
The simplest way to understand ChatGPT's Work is to separate a question from a job.
"Give me ideas for a podcast interview" belongs in Chat. "Review this guest brief, research the company's recent announcements, compare its claims with independent reporting, then prepare eight interview questions" belongs in Work.
This distinction matters because Work consumes usage according to how much effort a task requires. A long research assignment with several files, websites, and outputs will use more of your Plus allowance than a small request. OpenAI says Plus includes expanded access to ChatGPT Work, but limits still apply.
Throwing every request into Work would be a quick way to burn through capacity without gaining much in return. I would use Chat to think aloud, test an argument, or rewrite a paragraph. Work gets the heavier jobs where several steps must occur in the right order. Be deliberate.
Start with Work you already understand
OpenAI's own advice is sensible here. Start by giving Work a task you know well. That allows you to judge the result properly. If I ask it to prepare for a podcast interview, I know what good research looks like, which questions sound natural, and where it has swallowed a company claim without checking it. I can see the weak spots quickly.
Starting with an unfamiliar task removes that safety net. A polished spreadsheet can contain the wrong formula. A smart-looking report may rely on weak sources. A presentation can be beautifully made while saying very little. AI has always been capable of producing confident rubbish. Work gives that rubbish better formatting.
For the first few jobs, stay close. Watch how it interprets the request, inspect the sources, and stop it when it heads in the wrong direction. You can answer questions or change course while it is working. That intervention is part of the process, not evidence that the tool has failed.
Describe the finished result.
A normal ChatGPT prompt often asks for an answer. A useful Work request describes what should exist when the job is finished.
"Research this subject" leaves too much open to interpretation. Tell it who the Work is for, what files it should use, which sources are acceptable, how recent the evidence must be, and what it should create at the end.
For example, I might ask it to review an interview transcript, identify the strongest argument, check any statistics against sources, and produce an article plan written for business technology readers. I would also tell it to flag anything that cannot be verified.
That final instruction is easy to miss. Make uncertainty visible. Ask Work to show missing information, conflicting evidence, and assumptions rather than quietly smoothing over them. You want a useful colleague, not an eager intern pretending everything is fine. The prompt can be short. It needs to be specific.
Give it the raw material.
ChatGPT's Work becomes far better when it has real context. Upload the brief, spreadsheet, transcript, notes, and any previous articles or house styles it needs. If there is a format you want copied, provide an example.
Don't spend twenty minutes explaining the layout of a weekly report when you can attach last week's report and say, "Use this structure, update the figures, and show me what changed."
Projects are useful here. A Project can hold related chats, files, and instructions together, while project memory keeps the model focused on material from that space. OpenAI says this creates a self-contained context, which helps when Work runs over several sessions.
I would create separate Projects for each publication, podcast, or client. Put the relevant tone, audience, examples, and working files inside each one. Mixing everything invites bleed. A cybersecurity article should not suddenly sound like a podcast sponsor message because both were dumped into the same space. Keep the boundaries clean.
Connect apps with some restraint.
Plugins let ChatGPT pull information from services such as Gmail, Google Drive, Slack, Teams, SharePoint, calendars, and project tools. You can also point it toward a named plugin by typing the @ symbol followed by the app name. This is where Work becomes genuinely useful.
Work can gather material that would otherwise be scattered across inboxes, folders, and messages, then build something from it. Conference preparation is an obvious example. Work could review a calendar, speaker information, company research, and previous notes before creating an interview schedule or briefing document. But access should follow the job.
Connecting every account because you might need it someday creates more exposure, more clutter, plus a greater chance that irrelevant material finds its way into the result. So only connect what you need. Remove what you no longer use. Check which service will receive data before approving an action. Most importantly, please read the screen before clicking yes.
Use the desktop when a task spans multiple apps.
The web version is useful for cloud-based Work. The desktop app makes sense when the job involves local files, installed applications, or repeated movement between tools.
ChatGPT Work can use its built-in browser for research and web tasks. Computer Use can also click, type, and move files across applications while the job runs. That opens the door to jobs such as updating a tracker from downloaded reports, organizing research files, or moving information between a website and a document.
I would begin with low-risk admin. Rename a batch of clearly identified files. Extract figures from reports into a fresh spreadsheet. Build a draft presentation from approved source material. Watch what it does before allowing it anywhere near a live customer system. The agent can move faster than you. But remember it can also make mistakes faster.
Turn repeated Work into scheduled Work.
Scheduled Tasks may be the most useful part of Work for people who publish, report on, or manage recurring workflows. A task can run once, at a set time, or when a condition changes.
I could ask it to check selected technology news sources each morning, prepare a source-linked briefing, and highlight developments connected to my current interviews. A company might use it to refresh a weekly meeting document from new Slack messages. A sales team could monitor account activity, while a marketing team updates a campaign report when new figures arrive.
Start with a manual version. Run it several times, correct the weak points, then schedule it. Automating a poor process creates a poor result on time, every week. Reliable beats clever here.
Ask for evidence, then inspect it.
Work can complete a larger share of the process, but responsibility still sits with the person using it. That becomes especially obvious in journalism, finance, legal Work, or anything involving customers.
Ask for direct links to sources. Tell it to separate confirmed facts from company statements. Make it identify publication dates, conflicting numbers, and evidence that may now be old. Then open the sources.
A citation proves that a page exists. It does not prove that the page supports the sentence beside it. This is also where human judgment earns its keep.
Work can collect and compare. You decide whether the argument is fair, whether a quote has lost its context, and whether the finished piece deserves to leave the building.
Check your data settings before using client material.
A Plus account is an individual plan. OpenAI says content from personal services such as ChatGPT and Codex may be used to train its models unless the user opts out.
You can change this under Settings, Data Controls, then switch off "Improve the model for everyone." New conversations will remain in your history but will no longer be used for model training. OpenAI explains the setting here.
Temporary Chat offers another option for isolated conversations. Those chats do not appear in history, create memories or train the models, although OpenAI may retain a copy for up to 30 days for safety purposes. But it still doesn't permit you to upload confidential client information.
Company policy, contracts, and common sense still apply. Some work belongs in ChatGPT Business or Enterprise, where business content is not used for training by default. Some material should stay out of any AI service.
Treat the first result as Work in progress.
Remember, the first output is rarely the best one. So, review and challenge it. Ask what was missed, where the evidence is weak, and which assumptions shaped the result. If a section feels generic, say so. If the tone sounds like a press release, point to the offending paragraph. Feed those corrections back while the context is still active.
I often get better results by explaining why something is wrong instead of merely saying I dislike it. "This repeats the company's message without testing the claim" gives the model a usable correction. "Make it better" does very little.
ChatGPT can carry out bigger jobs than the chatbot most people first met. The tradeoff is that we now have to manage the Work, the permissions, and the finished result with greater care.
Used lazily, it produces larger piles of plausible content. Used well, it can remove hours of research, sorting, and production while leaving the judgment where it belongs. With us.