What Is ChatGPT Work? Features, Pricing and Access

ChatGPT Work is OpenAI’s agent for jobs that require several steps, multiple sources, and a finished file at the end. It can gather context from approved apps, plan the work, stay with a project for hours, and return editable documents, spreadsheets, presentations, reports, and shareable web pages it calls Sites. Access is included with eligible ChatGPT plans rather than sold as a separate subscription.

ChatGPT used to be something you asked. Work is closer to something you brief. A normal chat is still the quickest route to an answer, an idea or a rough draft. Work exists for the moment; the answer has to become something useful: a board report, a campaign plan, a spreadsheet, a deck, or a web app that somebody can review and use.

That shift matters for marketing and business teams, and what earns its keep is the coordination rather than the prose. Plenty of tools can write words. Far fewer can keep a goal in view across half a dozen sources, a spreadsheet and a slide deck, then hand back work that hangs together. The flip side needs saying too. The longer it runs and the more it can touch, the more the boring questions matter. Who granted access to what? How good were the source files? Whose name goes on the sign-off?

The launch came on 9 July 2026. OpenAI reshuffled its desktop line-up the same day, so Chat, Work and Codex now share one redesigned app on Windows and macOS. Its announcement fills in the engineering: Codex agent technology underneath, with the new GPT-5.6 models thinking.

The short version first.

ChatGPT Work: key features, access and pricing

Availability, models and usage limits may change as OpenAI continues the rollout.

What exactly is ChatGPT Work?

ChatGPT Work is an autonomous agent inside ChatGPT that gathers context across a user’s connected apps and files, breaks a goal into smaller steps, and works through complex projects independently. Unlike the standard Chat experience, which is designed for immediate and iterative assistance, ChatGPT Work is intended for longer projects involving multiple steps, sources, tools, and deliverables.

It has Codex technology built in, the coding agent more than five million people already use every week, and runs on the GPT-5.6 model family released the same day: Sol is the flagship model for demanding work, Terra balances capability and cost for everyday tasks, and Luna is the fastest and most affordable option. Before it starts, you see a plan you can review and adjust, and you decide what it can access, when it checks in, and which actions need your approval. OpenAI’s launch announcement has the full rundown.

What does ChatGPT Work actually do?

Mostly, it turns source material into usable files without losing the thread between them. Ask ordinary ChatGPT for a campaign idea, and you get an answer in the conversation. Ask Work to prepare a campaign launch, and it can hold the same research, brand guidance and decisions steady across a brief, a content calendar and a deck, so the three files agree with each other.

Along the way, it behaves less like a text box and more like a contractor. It reads whatever you have connected, works out what is missing, asks when a gap matters and carries on through the steps, sometimes for hours. Scheduled Tasks extend this to recurring jobs: a weekly meeting pack that refreshes itself, or a report that assembles whenever new project updates land.

ChatGPT Work vs standard ChatGPT

Chat has not been replaced, and for a quick explanation or a draft you plan to shape yourself, it remains the better tool. Work earns its keep when the job has a clear end state and several things must happen before you get there.

How to use ChatGPT Work

The mechanics are simple. The quality of the result depends almost entirely on the brief.

1. Switch from Chat to Work. On desktop and web, the Chat and Work toggle sits at the top centre of the screen. On mobile, tap the heading at the top and choose Work from the dropdown. Desktop users should also check the menu in the top-left corner is set to ChatGPT rather than Codex.
2. Start in the right place. A new conversation works, but choosing an existing Project brings its files, instructions, and earlier decisions with it.
3. Give it the material. Upload the relevant files and connect the apps this assignment needs through Plugins in the message box. Good context helps; indiscriminate access just widens the risk.
4. Describe the result, not the task. Say what you want produced, who it is for and how it will be used. Deadlines, tone of voice, branding, file format, formulas, what must not change. The lot.
5. Set the review criteria. Ask for claims to be sourced, assumptions to be marked and gaps to be raised rather than quietly filled in.
6. Check the proposed plan. For anything involved, it will ask clarifying questions and set out its steps. Correcting the route here is far cheaper than rebuilding files later.
7. Monitor, approve, then read the thing. The Ask for approval setting in the message box governs how much it checks in, and it will seek confirmation before actions that touch connected services or would be hard to reverse. Review the output properly before it goes anywhere near a client. A polished file can still contain a polished mistake.

One small habit worth forming: glance at the model picker before you send. It shows the model and its effort setting, and a big job on Sol at a high effort level does the deepest work while draining the most allowance.

Here is an example of a very concise brief that can be used as a starting point:

Prepare a 12-week UK launch plan for [product]. Use the attached customer research, proposition document and brand guide. Deliver a campaign brief in Word, a content calendar in Excel and a ten-slide internal deck. Keep every approved product claim unchanged, flag anything that needs evidence, write in UK English and stop for approval before doing anything outside ChatGPT.

OpenAI’s Work and Codex guide covers the mechanics in more detail.

What changed on launch day

On web and mobile, access opened with Pro, Enterprise, Edu, Plus and Business followed within days. On desktop, the redesigned app replaces the separate Codex application and covers every plan, including Free, while the old app lives on as ChatGPT Classic.

Two quieter changes arrived alongside it. Sites went into public beta, which lets a piece of work be published as a shareable web app with its own URL. And OpenAI confirmed it was winding down Atlas, its standalone browser, folding what it learned into ChatGPT itself and a Chrome extension. That wind-down is now complete: Atlas closed on 9 August 2026, ten months after it launched.

Who can use ChatGPT Work in the UK?

ChatGPT Work is available in the UK on every surface OpenAI offers, and what you get follows the plan you already have:

• Free and Go: a small amount of usage through the desktop app only
• Plus: access across desktop, web and mobile
• Pro: bigger allowances and the most capable model settings
• Business, Enterprise and Edu: access governed by workspace administrators, who also get spend controls through the Admin Console

One change since launch sits just outside Work itself but affects the same plans: since 6 August, free ChatGPT accounts run on GPT-5.6 Luna by default with unlimited text chats, which we cover in our GPT-5.6 Sol guide.

Does ChatGPT Work have a separate price?

No. It is not sold as a subscription or an add-on; it comes included with eligible plans, and the real cost sits in the metering. Every task draws on your plan’s usage allowance, measured the same way as Codex, so what a job consumes depends on the model, the amount of context, and how long it runs. A simple document costs little. A project that reads five files, queries your CRM and produces three outputs costs considerably more.

Once the included usage runs out, individuals on Plus and Pro can buy extra credits, and business workspaces on flexible pricing can buy shared ones. A caveat for anyone budgeting: OpenAI’s published usage examples mostly come from Codex coding tasks, so treat them as a rough guide rather than a forecast for Work projects. The ChatGPT pricing page has the current plan comparison.

ChatGPT Work vs Claude Cowork: key differences

Put the two side by side, and the differences sit in access and maturity rather than ambition. ChatGPT Work arrived everywhere at once: desktop on every plan, web and mobile from Plus upwards. Claude Cowork has the longer track record, running on desktop since its January research preview, but it remains a paid-plan product, with its web and mobile beta still rolling out from the Max plan.

The timing was no accident, either. Cowork went generally available on 9 April and reached web and mobile on 7 July, two days before OpenAI’s launch.

Comparison as of 10 August 2026:

Strip the branding and the two ask the same things of a business: an outcome worth delegating, decent source material and firm rules about what needs a human. Neither vendor disagrees on that last point. The differences are in the plumbing.

Anthropic’s own research explains why both firms are fighting over the same ground. Of the 1.2 million Cowork sessions it sampled in May, business process and operations accounted for 33 per cent, content creation another 16 per cent, with software development under 9 per cent. The agents built by coding companies are being used to do ordinary office work.

Neither product can be judged from a spec sheet, and the deciding factor for most businesses will be unglamorous: which one already connects cleanly to your files, applications and approval process. Pilot one tool on one workflow you know well, measure what that workflow costs you today, and compare the result rather than the launch claims.

Where ChatGPT Work fits in a marketing or business team

This launch fits a pattern we watch closely through our answer engine optimisation work: the tools are now built to finish the job rather than describe it. And the question we ask clients about being found in AI search now applies inside the business too. Is your content, data and process structured well enough for an AI system to use?

That is the part the tech coverage keeps missing. An agent can only build a decent pipeline report or campaign analysis if your files, CRM records and brand guidelines are in order. Point it at a mess, and it will produce confident rubbish. Tidy naming conventions and a single source of truth for brand assets have quietly become AI infrastructure.

What that looks like role by role:

The launch examples lean heavily towards marketing and commercial teams. Zapier’s head of enterprise marketing built a lead review system that analyses thousands of inbound leads a month across CRM and email, surfacing what OpenAI reports as seven figures in potential sales that follow-ups had missed. Virgin Atlantic’s digital team used it for competitor airline analysis, cutting cycles from weeks to hours. RingCentral automated its monthly launch checks across release plans and Jira, letting one efficiency manager support fifty product managers rather than one. All of these come through OpenAI’s own launch material, so read them as direction of travel rather than independent benchmarks.

What happened when we tested it

We ran our own test on 19 July, on a Pro plan through the desktop app with Sol set to Extra High. The brief: a fact-checked statistics pack on UK digital marketing, delivered as a 1,500-word report, an Excel register and a six-slide internal deck, with an exact source URL for every figure and firm instructions not to estimate, average or invent. It asked one approval question, proposed its plan immediately and then worked unattended for 32 minutes, using about three per cent of the week’s usage allowance.

It came back with 59 statistics drawn from five sources published by Ofcom, IAB UK, the DMA and DataReportal. We checked every one of the 59 against its cited source, and every one held up. Better still, it refused the bait: the universal email ROI multiplier and the “nobody scrolls past page one” figure were listed as could not verify rather than quoted, which is more discipline than plenty of human marketing content manages. The register, report, and deck agreed with each other exactly. The faults we found were cosmetic: a source name shortened inconsistently on one slide and a miscount in its own closing summary, which called five sources five publishers.

Doing this by hand, to that sourcing standard, we would budget a full working day, call it eight hours: two Ofcom reports to read, one of them long, plus the IAB, DMA and DataReportal releases, then the register, the report, the deck and the cross-checking. Reviewing the finished pack took around 45 minutes. That is a day’s deliverable for under an hour of human attention, most of it the review that nobody should skip. One run is a sample of one, and an easier brief to police than most. But on this evidence, the honest conclusion is the one in our verdict: a capable junior, and quick.

Budget deserves a thought as well. A team that points Sol, the dearest model, at every small job can drain a month’s allowance in week one, before anyone has asked whether the output justified it.

Is ChatGPT Work safe for company data?

It can be used responsibly with company information, but the safety lives in the configuration rather than the product. Work reaches connected services through app permissions you grant, asks approval before sensitive or hard-to-reverse actions, and on desktop can use approved local files. That convenience makes permissions part of the brief itself, and someone senior should own them.

Enterprise and Edu administrators can restrict connected tools, company context and permitted actions centrally, and a Compliance API gives them visibility of conversations and actions. OpenAI states that data from Business, Enterprise and Edu workspaces is not used to train its models by default; its business data policy sets this out. Personal accounts carry different data controls, so a sole trader on Plus should not assume the same protections as an approved business workspace. Third-party apps connected to ChatGPT also keep their own terms and retention practices, which deserve the same review.

Before letting an agent loose on business data, it is worth confirming a few things:

• Your organisation has an actual policy on agent tools and company data, rather than leaving it to individual judgement
• Someone has documented which apps, folders and data types the agent may reach, particularly open Slack workspaces and shared drives
• Access is granted per assignment at the minimum level needed, never a whole workspace by default
• External messages, publishing and data changes always need human approval
• Generated documents and decks are treated as drafts until a named person has checked them
• There is a record of the brief, the sources used and who signed off before anything went external

Our verdict

ChatGPT Work is at its best on structured, multi-source tasks where the expected output is clear, and a person reviews the result. It is a poor fit for undocumented workflows, unsupervised external communication, or decisions that need accountable professional judgement.

In other words, it is a very capable junior. It still needs a manager.

The sensible starting point is small. Take one task the team already understands, record what it costs in hours today, then run the same job through Work for a fortnight. Measure the time saved after review, count the corrections, and note where the agent needed better data or a clearer rule. That will tell you more about the business case than any product demo.