ChatGPT Work: OpenAI's Agent That Ships Finished Work

ChatGPT Work launched on July 9, 2026 — OpenAI's new agent inside ChatGPT that takes an outcome, gathers information across your apps and workflows, and stays with complex projects for hours by breaking them into smaller steps and completing them independently. The output isn't chat. It's finished materials: sheets, slides, docs, and shareable web apps.

The launch lands as a bundle of structural moves. The Codex app is merging into a new ChatGPT desktop app that puts Chat, Work, and Codex on every plan, including Free. The standalone Atlas browser begins sunsetting, its agentic-browsing learnings absorbed into ChatGPT itself. And the whole thing runs on GPT-5.6, which went generally available the same day. This is the moment the pieces OpenAI has been assembling all year — the team merge, Codex Sites, desktop computer use, the plugin directory — become one product with one name.

This guide covers what actually launched, the desktop-app merge, what the Atlas sunset signals, how connected tools become finished artifacts, what named early testers did with it, the staged rollout and the usage-metered pricing model, and whether it's safe to point at client data. Everything below is sourced from OpenAI's launch post and product page, both published today.

OpenAI describes ChatGPT Work 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 — sheets, slides, docs, and web apps — and stays with complex projects for hours by breaking them into smaller steps and completing them independently. The critical line in the announcement: "With Codex technology built-in, ChatGPT can now move beyond answering questions to getting real work done across web, mobile, and desktop."

The engine is GPT-5.6, which went generally available the same day — we cover the model itself in our GPT-5.6 GA analysis. Don't conflate the two: GPT-5.6 is the model; ChatGPT Work is the agent product built on top of it. Per OpenAI's product page, you give it an outcome and it "can navigate ambiguity, adapt as work unfolds, and deliver polished outputs with less prompting."

Why productize now? The adoption numbers OpenAI cites tell the story: Codex — originally a developer tool — had quietly become a general-purpose work agent.

OpenAI also says that "nearly 100% of teams inside OpenAI, including finance and sales, now use ChatGPT Work and Codex" — a self-reported figure, but a telling one about how the company sees the product's scope. None of this arrived from nowhere. When OpenAI merged its ChatGPT and Codex teams earlier this year, we read the org chart as a product roadmap — a unified assistant-plus-agent was the obvious destination. Today it has a name.

The structural news sits in one sentence of the launch post: "Starting today, the Codex app is merging with the new ChatGPT desktop app." Chat, Work, and Codex now live in a single app, available on every plan including Free. The existing ChatGPT desktop app is renamed ChatGPT Classic. Developers lose nothing — they can keep Codex as their default view and even keep the Codex logo as the app icon.

Codex itself gains capabilities in the merge: inline editing within diffs, PR review in the side panel, faster computer use powered by GPT-5.6, and multi-repo projects. But the bigger shift is what the desktop app means for non-developers. Desktop is the power surface: the agent can reach local files and apps, drive a built-in multi-tab browser for agentic workflows, and use Computer Use — it clicks, types, and moves files on your behalf — for one-time jobs or Scheduled Tasks.

If the desktop capabilities sound familiar, that's because they shipped first as Codex features — we covered the built-in browser, computer use, and the plugin system in depth in our Codex desktop and computer-use guide. What's new today isn't the capabilities; it's that they stop being a developer product's features and become the default work surface for everyone on the plan, under one roof.

Buried in the launch post is a genuine strategy reversal: "We'll begin sunsetting the standalone Atlas browser." The agentic-browser learnings move into ChatGPT itself — the desktop app's built-in browser — plus an updated Chrome sidebar extension. Note the phrasing: this is the announced start of a sunset, not a shutdown. Atlas users are being transitioned, not cut off today.

When Atlas launched, we argued in our Atlas strategy analysis that the browser was fundamentally a distribution play — a way to put OpenAI's agent where the work happens. Today's move completes that thesis in an unexpected direction: distribution didn't need a standalone browser after all. Once the assistant itself ships with a built-in browser on every desktop, the browser-as-product becomes redundant with the assistant-as-product.

The category read is worth sitting with. If the best-resourced AI lab in the world couldn't justify maintaining a standalone AI browser alongside its assistant, the standalone AI browser as a category now has a serious existence question. The durable pattern appears to be agents that reach into browsing as a capability — not browsers that bolt on an agent.

The pipeline runs in four moves. First, context: more than 1,400 plugins connect the tools you already use, you can @-mention an app to pull its context into a task, and ChatGPT auto-suggests relevant plugins as work unfolds. Second, planning: in Plan mode, ChatGPT gathers context, asks clarifying questions, and produces a step-by-step plan you approve or adjust before any work begins. Third, execution — including on a schedule. Scheduled Tasks run once, on a schedule, when an event occurs, or continuously as monitors: review Slack updates weekly and refresh a meeting agenda, turn a stream of customer feedback into prioritized product ideas, update a deck when new feedback arrives by email.

Fourth, the artifact itself. Beyond sheets, slides, and docs, Sites — now in public beta inside ChatGPT — turns work into interactive websites and web apps shared via URL: live dashboards, project trackers, launch calendars, prototypes, internal portals, interactive reports. ChatGPT can keep them updated as the underlying information changes. We covered the Sites concept when it launched under the Codex banner in our Codex Sites team guide — what's new today is that it lives in ChatGPT proper, in public beta, as the agent's output format of choice for anything a static doc can't hold.

OpenAI's own worked example is agency-shaped: turn customer research into a campaign brief, generate the marketing assets, and adapt those assets for different markets — carrying context through every step in a single request. That is a brief-to-assets pipeline, the connective tissue of agency delivery. Throughout, you stay in control: you decide what it can access, when it checks in, and what needs approval, and you can follow progress, answer its questions, and change direction mid-run.

The most useful part of the launch materials is the tester roster — not because vendor-arranged testimonials are neutral evidence (they aren't; every example below comes from OpenAI's early-access program and should be read that way), but because the workflows named are strikingly concrete and strikingly marketing-ops. Lead triage. Competitor benchmarking. Event reporting. Launch readiness. These are the exact workflows agencies and in-house teams staff people against.

The Zapier example rewards a closer look because it's the shape of work most revenue teams recognize. A single inbound lead took Ferrante's team 35–45 minutes to properly inspect — pulling CRM history from HubSpot, call recordings from Gong, and the email thread. Multiply by lead volume and the honest answer is that most leads never got that inspection. The agent-built QA/QC system traces each lead's journey and surfaces drop-offs, which is what turned unexamined volume into the pipeline figure she cites. If your CRM has the same dark corners, this is the workflow class our CRM automation engagements are built around — instrument the pipeline first, then let agents work it.

So what's the day-one play? OpenAI's implicit advice — visible in every tester story — is to point the agent at one workflow you already know intimately, not to reorganize around it. Here's how we'd frame the choice for a marketing or revenue team:

Whichever you pick, run it as an experiment, not a rollout: Plan mode on, approvals on, and your current hours-per-cycle written down before the agent touches anything. The tester numbers above are self-reported under a vendor's early-access program — your number is the only one that should drive a staffing or budget decision.

Availability is staged, and OpenAI's own materials differ on the desktop detail — the launch post says the new desktop app is global today, while the product page's fine print stages Windows. The reliable read: macOS desktop is live today on all plans; Windows is rolling out over the next few days.

The pricing model deserves more attention than the rollout dates. ChatGPT Work is not a flat feature of your subscription — OpenAI is explicit that it "is designed for longer, more involved work than a typical chat request, so usage works differently," and that "ChatGPT Work follows the same usage structure as Codex." Usage varies with the amount of work a task requires. Enterprise and Edu admins get spend controls in the Admin Console: workspace defaults, group limits, individual overrides, and a request-for-credits review flow.

The strategic pattern is one we've tracked all year: AI vendors are converging on hybrid pricing where the subscription buys an allowance and ambitious agentic work meters against it. If you're planning team budgets around this launch, our breakdown of subscriptions versus usage credits across OpenAI and Anthropic is the companion read — the failure mode it warns about, a team burning its month's allowance in the first week of enthusiastic agent use, is precisely the risk profile of a product designed to run for hours.

For any team handling client data, the launch's most consequential section is the quietest one. ChatGPT Work is built on ChatGPT Enterprise's security, privacy, compliance, and workspace-management foundation. Admins centrally manage who has access to the agent, what context it can use, which tools it connects to, and what actions it can take. A Compliance API provides visibility into conversations and actions at scale — the audit-trail requirement that usually blocks agent adoption in regulated or client-confidential environments. On desktop, the governance model builds on Codex's enterprise controls, including network-access policies for agents.

The new mechanism is auto-review: advanced models review important actions involving connected tools and APIs before they happen. OpenAI's headline result needs its qualifier kept attached.

That last point is where we'd anchor any deployment. The approval architecture — Plan mode before work starts, check-ins you configure, actions gated on your sign-off — is not friction to be toggled off once trust builds. It's the mechanism by which senior judgment stays in the loop while the agent does the volume. That's the operating model we build for clients in our AI transformation engagements: agents run the workflow; accountable humans own the decisions the workflow feeds. ChatGPT Work shipping with that architecture as a first-class feature, rather than an afterthought, is one of the stronger signals in the launch.