QWE AI Academy You give ChatGPT a sentence like “pull last month’s support tickets from Zendesk, group them by product area, and drop a summary deck in my Drive.” You walk away. A while later there’s a finished Google Slides deck sitting where you asked for it – no chat back-and-forth, no copy-paste between tabs. That’s the pitch for ChatGPT Work, and as of July 9, 2026 it’s actually shipping.
This guide walks backward from that end-state: what the tool is, how to set it up in about ten minutes, one advanced move worth learning early, and the pricing detail nobody’s warning you about yet.
The scenario: what “finished work” actually means
Regular ChatGPT answers questions. ChatGPT Work produces artifacts. OpenAI describes it as “an agent in ChatGPT that helps you take on more ambitious tasks” – it gathers information across your apps and workflows to create finished materials like sheets, slides, docs, and web apps, and stays with complex projects for hours by breaking them into smaller steps.
A useful mental model: think of it less as a chatbot and more as a junior teammate with a laptop. It reads your Slack, opens your Drive, drafts the thing, and hands it back.
What just launched (and what didn’t)
July 9, 2026. The Codex app merges into the new ChatGPT desktop app, and OpenAI begins sunsetting the standalone Atlas browser. Under the hood, ChatGPT Work runs on the just-released GPT-5.6 – three model tiers, each with a different price point and capability ceiling.
The tier breakdown (per OpenAI’s announcement):
| Model | Role | API price (per 1M tokens) |
|---|---|---|
| Sol | Flagship, hardest tasks | $5 in / $30 out |
| Terra | Balanced everyday | $2.50 in / $15 out |
| Luna | Fast, cost-efficient | $1 in / $6 out |
Community reaction on Reddit and X so far has been split between “finally a real agent” and “this is just Codex with a new coat of paint.” There’s truth in both. Many of the featured functions – Scheduled Tasks, plugins, Computer Use – already existed in ChatGPT or Codex. The difference is packaging, not primitives.
Setup in about 10 minutes
The fastest path depends on your plan. There’s a rollout asymmetry worth knowing before you start.
- Pick your surface. Web and mobile start with Pro, Enterprise, and Edu – Plus and Business follow in the coming days. Through the desktop app for Mac and Windows, Work is available immediately for all plans, including the free tier. If you’re on Free or Plus and can’t wait, install the desktop app.
- Grab the desktop app. The old ChatGPT desktop app is renamed ChatGPT Classic; the new one contains Chat, Work, and Codex together. Existing Codex app users can update their current install to get it. One catch, as of July 10, 2026: the Windows link on the OpenAI download page still leads to the Codex app – update through the Codex installer if you’re on Windows rather than following the official download flow.
- Open Plugins. On the web, select Work and open Plugins. In the desktop app, select Work or Codex and open Plugins. (Full walkthrough in the official plugin docs.)
- Connect two or three apps. Don’t connect everything at once – start with apps you actually pull from daily. At launch the directory includes Google Drive, SharePoint, Slack, Microsoft Teams, Gmail, Outlook, Salesforce, Adobe, Zoom, LinkedIn, GitHub, Canva, and Dropbox, among others.
- Test with an @-mention. Type “@” followed by the app name you want Work to pull from. Something like “@Gmail summarize threads from last week tagged ‘launch’ into bullet points” is a good first probe.
Pro tip: Before pointing Work at anything sensitive, turn on Lockdown Mode from Settings → Security. It’s an opt-in setting that restricts live browsing, deep research, agent mode, and file downloads – the trade-off is less reach for lower prompt-injection risk. Worth it for your first few runs. (Documented in OpenAI’s release notes.)
Scheduled Tasks with plugins – the move that actually changes things
Scheduled runs. Tell the agent to do something on a cadence; it hands you the result.
@Slack + @Gmail
Every Monday 8am:
- Pull all messages from #customer-feedback since last Monday
- Pull emails to support@ tagged "bug"
- Group by product area
- Draft a one-page summary in Docs, share link in DMThat’s not hypothetical syntax – it’s the shape of what the Scheduled Tasks feature does: automatically performing repetitive work on a set schedule or when specific events occur. Pair that with the Sites beta and it turns work or ideas into interactive sites or web apps shared via a URL – real-time dashboards, project tracking tools, prototypes, internal portals.
How valuable is that, really? Depends on whether your workflows are actually repetitive enough to justify setup time. For a one-off task, just prompt it. For something you’d otherwise do manually every Monday morning? The scheduled approach pays back within a week.
The pricing trap nobody’s talking about
Here’s the detail buried in the launch docs that will burn people. Work doesn’t have a clean “messages per day” cap like classic ChatGPT.
Shared pool. ChatGPT Work, Codex, ChatGPT for Excel, and Workspace Agents all draw from the same agent-based consumption bucket – and how much a task consumes depends on its size, complexity, and the model selected. Run a heavy Codex refactor in the morning, and the pool that would have powered your afternoon Work run is already drained.
Turns out the Codex pricing page lists included quotas and extra credits, but those examples are calibrated on coding tasks. OpenAI explicitly notes consumption for Work tasks may differ – so any budget you built from Codex numbers is a rough guess, not a plan.
Workaround: Enterprise and Edu admins get Spend Controls at the workspace, group, and individual user levels. On Plus and Pro, watch your usage panel for the first week before scheduling anything expensive.
What doesn’t work yet
Web access lags desktop for Free and Plus users right now. If a teammate says “I don’t see Work,” send them to the desktop app.
The critics have a point about the rebrand angle. Much of what Work does – long-horizon tasks, plugin connections, structured outputs – was already possible in Codex or Workspace Agents. If your existing setup works, don’t rip it out to chase the new branding.
Auto-Review exists (advanced models check important actions before they run), but treat it like a spellchecker, not an approver. Review anything that touches production data yourself.
FAQ
Do I need to uninstall ChatGPT Classic?
No. Keep both. Classic is your regular chat; the new desktop app is where Work and Codex live.
Which model should I actually use for Work tasks?
Default to Terra for most business tasks – it’s the balanced option and eats less of your consumption pool. Reach for Sol when the task involves multi-step reasoning across many files, like reconciling three quarters of financial reports or producing a real strategy doc. Luna is best for volume work: hundreds of short classifications, quick email drafts, batch renaming across a Drive folder. A common mistake is defaulting to Sol “just to be safe” – you’ll burn through quota on tasks Terra would have handled fine.
Is my company data safe to point at?
Depends on your plan and your admin’s settings. Enterprise and Edu workspaces have proper controls; personal accounts don’t. Turn on Lockdown Mode for your first runs and never let Work touch a system you couldn’t safely hand to a new contractor on day one.
Do this next: open the desktop app, connect only Google Drive (or your equivalent), and ask Work to summarize one specific folder into a one-page Doc. That single run will teach you more about how it actually thinks than another hour of reading will.