ChatGPT Work and Codex: OpenAI's Superapp on Every Plan

ChatGPT Work is an agent mode inside ChatGPT that takes a goal, pulls context from your connected apps and files, and returns finished sheets, slides, docs or web apps instead of chat replies. OpenAI launched it in July 2026 alongside GPT-5.6, merged the standalone Codex app into a new ChatGPT desktop app, and put Chat, Work and Codex in that desktop app on every plan, including Free. The strategic read: OpenAI is using its consumer reach to put agents in front of as many people as possible, while Anthropic has built its business on enterprise and developer usage. For you, the practical question is which recurring tasks are worth handing to Work, and how to keep it from quietly spending your allowance.

This guide covers what actually shipped, what each plan gets as of September 2026, where Work earns its keep, and where it doesn't. Every plan detail below comes from OpenAI's own launch post, help center and pricing documentation.

What did OpenAI actually launch?

Three things shipped together on July 9, 2026, the launch date given in the GPT-5.6 announcement on OpenAI's Developer Community forum.

• ChatGPT Work. In its launch post, OpenAI describes Work as "an agent in ChatGPT" that can "gather information across your apps and workflows to create finished materials like sheets, slides, docs, and web apps, and stay with complex projects for hours by breaking them into smaller steps." It runs in the cloud on web and mobile, and it can also run locally in the desktop app.
• A merged desktop app. The Codex app became the new ChatGPT desktop app. According to OpenAI's migration guide, existing Codex users just update and land in the new app with their chats and projects intact. The previous ChatGPT desktop app lives on as "ChatGPT Classic," which still gets model updates, but "new agent features may be available only in the new app."
• GPT-5.6. The model family that powers Work comes in three tiers: Sol (flagship), Terra (balanced) and Luna (cheapest and fastest). Per OpenAI's GPT-5.6 announcement, Free and Go users get Terra in Work and Codex, while paid plans can choose among all three.

The rest of the launch post fills in the mechanics. Plugins connect ChatGPT to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRMs and project trackers. Typing "@" plus an app name points Work at a specific source. Scheduled Tasks run a job once, on a schedule, when an event occurs, or as ongoing monitoring. Sites, in public beta, turns work into a shareable interactive page or web app. On desktop, Work also gets a built-in browser and Computer Use, which lets it click, type and move files across your apps. OpenAI also said it will begin sunsetting its standalone Atlas browser in favor of these capabilities.

Why did Codex end up inside ChatGPT?

OpenAI gave a usage reason. It says more than 5 million people use Codex every week, and more than 1 million of them use it for work outside software development. A coding agent was already doing non-coding jobs, so OpenAI moved the same engine into the product its general audience already opens. The help center describes the split in one line each: Chat is "for fast, conversational assistance," Work is "an agent designed for longer, multi-step work and finished deliverables," and Codex stays "dedicated to software development and technical work."

For developers, the merge is mostly additive. Codex remains a separate view with its own history. It is not selectable on web or mobile, though the mobile app's Remote tab can reach supported desktop Codex chats. The launch added inline editing within diffs, pull request review in a side panel, faster computer use, and multiple repositories in one project. If you are comparing agentic coding tools, our Claude Code review and our breakdown of AI coding tool pricing give the competitive context.

Is ChatGPT Work really free on every plan?

Yes, with conditions that matter. OpenAI's launch post says that in the ChatGPT desktop app, "Chat, Work, and Codex are available to users on every plan, including Free." On web and mobile, Work was rolled out to paid plans only: Pro, Enterprise and Edu first, then Plus and Business. The help center still frames web and mobile access as "eligible paid plans." Free users therefore reach Work through the desktop app, which launched on Mac and Windows. A Linux preview followed, announced on the OpenAI Developer Community forum on August 11, 2026, and OpenAI's desktop app documentation now lists Linux alongside macOS and Windows.

The bigger catch is usage. Work "follows the same usage structure as Codex," and OpenAI's Codex pricing page states that Work "uses the same pricing, credits, and usage limits as Codex." According to the help center, Codex, ChatGPT Work, ChatGPT for Excel and Workspace Agents draw on one shared allowance where your plan includes them. OpenAI describes the Free tier of that allowance as enough to "explore Codex capabilities on quick coding tasks," and Go as covering "lightweight coding tasks." So Free Work is a real product with a small budget. It is not a free agent running all day.

Plan	Where Work runs	GPT-5.6 models in Work	Notable limits (as of September 2026)
Free	Desktop app	Terra	Small shared allowance; no Sites; no event-triggered tasks
Go	Desktop app	Terra	Higher allowance than Free; no Sites; no event-triggered tasks
Plus	Web, mobile, desktop	Sol, Terra, Luna	Sites (outside the EEA, Switzerland and UK); event-triggered tasks; can buy extra credits
Pro	Web, mobile, desktop	Sol, Terra, Luna	5x or 20x Plus usage; ultra multi-agent mode in Work
Business	Web, mobile, desktop	Sol, Terra, Luna	Workspace admin controls over access, plugins and roles
Enterprise / Edu	Web, mobile, desktop	Sol, Terra, Luna	Spend controls; Compliance API (Enterprise); ultra in Work on Enterprise

Sources: OpenAI's launch post, GPT-5.6 announcement, help center and Codex pricing page. Sites and event-triggered tasks (Gmail, Slack and GitHub triggers) are excluded from Free and Go per the help center. Model menus change fast: OpenAI's help center now also lists GPT-6 Astra in Work and Codex for Plus (with limited usage), Pro, Business and Enterprise, so check the model picker rather than trusting any static table.

Why is OpenAI chasing seats while Anthropic chases usage?

The two companies describe their businesses differently, and the difference shows up in this launch. OpenAI now reports that an advertising-supported free tier "helps keep ChatGPT available to more than 1 billion weekly active users," and it calls advertising one pillar of a model that also includes consumer subscriptions, enterprise offerings and usage-based APIs. Putting an agent on the Free desktop app turns that audience into a pool of people who have tried delegating work. Some will outgrow the allowance and upgrade. Others will bring the habit into a workplace that buys Business or Enterprise seats.

Anthropic presents itself the other way around. In its Series G announcement in February 2026, it called Claude "the intelligence platform of choice for enterprises and developers." It said more than 500 customers were spending over $1 million a year on an annualized basis, and that Claude Code's run-rate revenue had passed $2.5 billion. By its Series H in May 2026, Anthropic said its run-rate revenue had crossed $47 billion. The growth story it tells is about businesses spending more on usage, not about how many people have an account.

The products are converging even as the business models differ. Anthropic's Cowork, launched in January 2026, brings Claude Code's capabilities to broader knowledge work, the same move OpenAI made with Work. What differs is who gets it first and how it is paid for. OpenAI leads with reach and a free entry point. Anthropic leads with enterprise contracts and API volume. For a team choosing tools, the result is that OpenAI makes it easy for individuals to adopt an agent before IT has decided anything. That is a benefit when you want to pilot quickly, and a governance problem if you don't know what data people are connecting.

What is ChatGPT Work actually good for?

OpenAI's own advice is the best starting rule: give Work "a task you already know well." Its examples are a month-end budget variance, a campaign brief built from source material, and prep for a sales meeting. The logic holds. You can only tell whether an agent did a task correctly if you know what correct looks like. Four patterns fit Work's design:

1. Template-bound deliverables. OpenAI claims GPT-5.6 is notably better at following templates and reference decks, including inferring a deck's layouts, typography and colors. If your team rebuilds the same weekly report or deck from new numbers, that is the strongest early use case, and it is easy to check.
2. Cross-app synthesis. Pulling threads from Slack, a shared drive and a CRM into one brief is tedious for a person and routine for an agent with plugins. Keep the output a draft you review, not a message it sends.
3. Monitoring on a schedule. OpenAI's examples include checking websites and dashboards each morning and summarizing what changed, or turning recurring customer-feedback themes into prioritized product ideas. These are low-risk because the output is a summary, not an action.
4. Quick internal tools with Sites. Dashboards, trackers and launch calendars that nobody would staff an engineer to build. Sites is a paid-plan feature in public beta and is unavailable in the EEA, Switzerland and the UK.

Where should you not use it yet?

Three limits deserve attention before you delegate anything that matters.

Data residency. The help center is explicit that when Work runs locally in the desktop app, "messages and task context may be stored in the cloud, even when work runs locally." Local execution is not the same as local data. Grant access to specific folders, not your whole drive.

Unbounded tasks. OpenAI says usage "varies with the amount of work required" and that similar-looking tasks can consume very different amounts. Its pricing page estimates that a Plus subscriber gets roughly 10 to 100 local GPT-5.6 Sol messages per five-hour window, versus 250 to 2,000 with Luna, and cloud chats on ChatGPT plans run on Sol. A vague, open-ended scheduled task is the fastest way to burn a week's allowance. Scope tasks tightly, pick the smallest model that works, and check the usage dashboard for the first couple of weeks.

Irreversible actions. OpenAI stresses that "you decide what it can access, when it should check in, and when it needs your approval before taking action." Use that control. Require approval for anything that sends, deletes, pays or publishes. The same principle applies to every autopilot-style agent, as we noted in our look at Microsoft Scout. On Business and Enterprise, admins can separately enable Work Cloud and Work Local by role and restrict plugins. For enterprise workspaces, OpenAI adds Auto-review, which uses its most advanced models to review important actions.

What should you do now?

• Individuals on Free or Go: install the new desktop app and test Work on one bounded, familiar task. Treat the allowance as a trial budget, not a work budget.
• Plus and Pro users: pick one recurring deliverable, build it once with Work, then turn it into a Scheduled Task with approval gates. Measure its usage before adding a second.
• Team leads and IT: assume people are already connecting Slack and Drive to a free agent. Decide which plugins are allowed, set Work Local and Work Cloud by role, and move serious use onto a managed workspace: Business adds admin controls over plugins and roles, and Enterprise adds spend controls and the Compliance API.

The bottom line: ChatGPT Work is the first agent most knowledge workers will meet, because OpenAI put it where they already are. Its value depends on how well you scope tasks, not on the model. The next thing to watch is whether OpenAI's free-tier bet converts into paid seats faster than Anthropic's enterprise-usage strategy compounds. Whatever the answer, the model picker and allowances will keep changing, so recheck the plan details above each quarter.

Frequently Asked Questions

What is the difference between ChatGPT Chat, Work and Codex?

Chat is ChatGPT's standard conversational mode for quick questions and brainstorming. Work is an agent for longer, multi-step tasks that end in a deliverable, such as a spreadsheet, presentation, report or Site. Codex is OpenAI's software-development agent, working with repositories, terminals and developer tools. Chat and Work appear together in ChatGPT, while Codex has its own view and history in the desktop app.

Can I use ChatGPT Work on the free plan?

Yes, through the ChatGPT desktop app on Mac or Windows (Linux is in preview), where OpenAI says Chat, Work and Codex are available on every plan, including Free. Work on web and mobile is limited to eligible paid plans. Free users run Work on GPT-5.6 Terra with a small shared usage allowance and cannot use Sites or event-triggered scheduled tasks, as of September 2026.

Does ChatGPT Work share limits with Codex?

Yes. OpenAI states that Work uses the same pricing, credits and usage limits as Codex. Codex, Work, ChatGPT for Excel and Workspace Agents draw from one shared allowance where your plan includes them. Complex or long-running tasks consume more, and Plus and Pro users can buy additional credits when they run out instead of upgrading plans.

What happened to the Codex app and the old ChatGPT desktop app?

Updating the Codex app turns it into the new ChatGPT desktop app, keeping existing Codex chats and projects. The previous ChatGPT desktop app is renamed ChatGPT Classic. It still receives model updates, bug fixes and security patches, but OpenAI says new agent features may be available only in the new app.