Introducing ChatGPT Work

On July 9, 2026, OpenAI shipped ChatGPT Work, an agent inside ChatGPT that takes action across connected apps and files, runs multi-hour projects by breaking them into steps, and produces finished sheets, slides, docs, and web apps rather than just answering questions. It runs on GPT-5.6, OpenAI’s new frontier model rolling out the same day, and inherits Codex technology as its execution engine.

OpenAI cites more than 5 million weekly Codex users, with more than 1 million now using Codex for non-coding work, as the evidence base for extending the same agent pattern to general knowledge work.

What shipped

The launch bundles several things at once.

• ChatGPT Work is rolling out on web and mobile, starting with Pro, Enterprise, and Edu, expanding to Plus and Business over the following days.
• The Codex desktop app is being merged into a new unified ChatGPT desktop app for Mac and Windows, where Chat, Work, and Codex are all available on every plan including Free. The existing desktop app is renamed ChatGPT Classic.
• The desktop app now includes a built-in browser and Computer Use, which lets the agent click, type, and move files across local apps.
• Sites in ChatGPT is launching in public beta to publish live dashboards, trackers, prototypes, and internal portals to a URL.
• Scheduled Tasks lets the agent run recurring or event-triggered jobs.
• A unified plugins directory connects Slack, Microsoft Teams, Google Drive, SharePoint, Gmail, Outlook, Salesforce, Adobe Acrobat, calendars, CRMs, and project trackers, invocable by prompt or by “@app” reference.

OpenAI also announced it will sunset the standalone Atlas browser, folding those learnings into the new desktop browser and an updated Chrome sidebar extension. Customer proof points come from Zapier, RingCentral, Virgin Atlantic, and NVIDIA, and OpenAI states that nearly 100% of teams inside OpenAI, including finance and sales, now use ChatGPT Work and Codex.

5M+weekly Codex users

1M+using Codex for non-coding work

~100%OpenAI internal teams on Work and Codex

Key details

Release date: July 9, 2026. Underlying model: GPT-5.6, rolling out the same day. Execution substrate: Codex technology, built in.

Rollout order: Pro, Enterprise, and Edu today on web and mobile; Plus and Business over the next few days. The desktop app is available globally today for Mac and Windows, with Chat, Work, and Codex on every plan including Free.

Customer proof points

• Zapier built a repeatable lead-review system tracing touchpoints across CRM, email, and other tools, generating a weekly executive dashboard that surfaced seven figures in potential sales.
• RingCentral turned monthly launch checks into a repeatable workflow reviewing release plans, Jira tasks, and go-to-market schedules, scaling support from one product manager to roughly 50.
• Virgin Atlantic compressed weeks of competitor passenger-experience analysis into hours for a five-year plan.
• NVIDIA automated GTC prep that had previously consumed about 40% of pre-event time and synthesized hundreds of session transcripts and customer-meeting notes afterward.

Internally, OpenAI says sales turned discovery calls into tailored proofs of concept in 24 hours versus weeks, and finance reduced month-end close and forecasting from days to hours.

Governance

The Compliance API is extended to ChatGPT Work. Auto-review gates important connected-tool and API actions. The Admin Console spend controls include workspace defaults, group limits, individual overrides, and a credit-request review flow.

Why it matters

This is the release where OpenAI stops pretending ChatGPT is a chatbot with tools bolted on and starts positioning it as the default place office work gets done. “Work” is now a top-level mode alongside Chat and Codex, not a feature buried in a menu. That is a bet that the agent surface, the model, the browser, and the desktop OS integration all belong under one product name, and that enterprises will consolidate onto whichever vendor can credibly deliver all four.

The competitive read is sharper than it looks. Microsoft 365 Copilot lives inside Office and, per OpenAI’s sibling announcement, now defaults to GPT-5.6. Anthropic’s Claude has agent capabilities and computer use but no first-party consumer scheduling layer at this scale. Google’s Gemini is embedded in Workspace but has not shipped a comparable “one agent that owns your project for hours” surface. What OpenAI is claiming here is the horizontal seat: the agent that sits above your SaaS stack and reaches down into any of it, whether you are a Microsoft shop, a Google shop, or both.

The Atlas sunset is the tell. OpenAI decided the browser was a means, not the product. The product is the agent, and it goes wherever the user already is.

The Codex merge matters for a second reason. The same execution engine hardened on a million-plus non-developer users is now the substrate for the general agent. That is a different maturity curve than a green-field agent launch, and it partially explains why OpenAI feels comfortable pushing Computer Use, multi-hour autonomy, and Scheduled Tasks in the same release.

What changed practically

In ChatGPT, Work becomes a distinct mode you can invoke. It accepts a goal, breaks it into steps, executes across connected apps, and produces artifacts. You can follow progress, redirect it, and approve sensitive actions. Plugins are invoked automatically or by typing ”@” followed by an app name.

1. 1Accepts a goal
2. 2Breaks it into steps
3. 3Executes across connected apps
4. 4Produces finished artifacts

Scheduled Tasks accepts one-off, recurring, or event-triggered runs. Example patterns include reviewing new Slack updates weekly to refresh a meeting agenda, checking dashboards each morning and sending a summary, monitoring customer feedback and turning themes into prioritized product ideas, and updating a presentation when new email feedback arrives.

Sites in ChatGPT, in public beta, produces shareable URLs for live dashboards, project trackers, launch calendars, prototypes, internal portals, and interactive reports. ChatGPT can update the underlying data over time, and users can test Sites inside ChatGPT before publishing.

On desktop, a built-in browser handles web research, tool use, and web-based file editing, with Google Workspace and Microsoft 365 files opening inside the app. Computer Use runs local tasks in the background across apps, tools, and browser, and can be embedded as a step inside a Scheduled Task. The Chrome extension is being updated to run ChatGPT in Chrome’s sidebar.

Codex, now inside the unified desktop app, gains inline editing within diffs, pull request review in the side panel, faster computer use powered by GPT-5.6, and multi-repository projects. Codex projects sync to the ChatGPT mobile app.

What to watch next

Three things to watch over the next quarter.

1. Whether the Codex-usage argument holds as a maturity proxy. Coding has tight feedback loops and testable output. Slide decks, campaign briefs, and month-end reconciliations do not. Expect the first public postmortems within sixty days, focused on how often the agent produces plausible-looking but wrong artifacts when it runs unattended for hours.
2. The Atlas sunset as a category signal. It suggests OpenAI has stopped believing in standalone AI browsers as a product. Watch whether other vendors investing in that surface adjust, and whether the Chrome sidebar plus desktop browser captures enough behavior to justify the retreat.
3. Governance will decide enterprise pace. Auto-review, Compliance API coverage of Work actions, and the spend-controls Admin Console determine whether large regulated buyers greenlight Work for more than pilots.

Expect competitive responses to hit the same bundle: agent mode plus scheduling plus artifact publishing plus a browser surface plus enterprise governance, delivered as one story rather than five.

Open questions

Several things the announcement does not resolve.

• Usage economics are described only as “follows the same usage structure as Codex,” with the caveat that complex tasks consume more of the plan’s allowance. A real unknown for anyone forecasting adoption cost at team scale.
• The Atlas transition path is named but not detailed. What happens to Atlas-specific workflows, saved contexts, and team deployments is not addressed.
• Auto-review depth is not quantified. Which actions trigger review, latency impact, false-positive behavior, and admin customization are all open.
• Scheduled Tasks ships without concurrency, reliability, or SLA language, so its operational envelope is unclear as a cron replacement.
• Computer Use on desktop has real blast radius, but the source does not detail sandboxing, rollback, or what happens on a destructive local change during an unattended task.
• Sites in public beta ships without stated limits on concurrency, hosting duration, custom domains, or access controls beyond “share via URL.”
• The Plus and Business rollout is “over the next few days” without a firm date, and feature parity across tiers is not confirmed.

Source: openai.com