ChatGPT Work Runs Your Job for Hours. The Codex App and Atlas Browser Just Died to Make Room.

TL;DR

On July 9, OpenAI launched ChatGPT Work, an agent inside ChatGPT that takes an outcome, pulls context from your connected apps, breaks the job into steps, and stays on it for hours until it hands back finished spreadsheets, decks, documents, or working web apps. It runs on GPT-5.6, which went generally available the same day. The quieter, arguably bigger story is the demolition underneath: the standalone Codex app is merging into one unified ChatGPT desktop app, the old desktop app is being renamed ChatGPT Classic, and the Atlas browser gets deprecated on August 9. And agent time is now usage-billed, not flat-rate.

What ChatGPT Work actually does

The pitch is outcome-in, artifact-out. You hand it a goal, it @-mentions its way through your connected tools (Slack, Microsoft Teams, Gmail, Google Drive, SharePoint, Salesforce, calendars, project trackers), gathers what it needs, and works the task in the background, checking in when it hits a decision it should not make alone. OpenAI says it will stay with a complex project for hours, decomposing it into smaller steps and finishing them independently.

The output is not a wall of chat text. Work ships finished files: sheets, slides, docs, dashboards, and, through a public beta called Sites, interactive web apps that live at a shareable URL and refresh when the underlying data changes. There is also a built-in browser with computer-use capability, so the agent can navigate sites and operate apps rather than just read them.

Rollout: web and mobile for Pro, Enterprise, and Edu on day one, with Plus and Business following over the next few days. The new desktop app is out for Mac and Windows, and every plan, including Free, gets Chat, Work, and Codex surfaces in it.

Three apps walk in, one walks out

For builders, the consolidation is the part worth staring at. OpenAI had sprawled into a small software suite: a ChatGPT app, a separate Codex app for agentic coding, and Atlas, a whole standalone AI browser. As of July 9 that is one desktop app with three surfaces. The Codex app merges in, the legacy desktop app becomes ChatGPT Classic, and Atlas gets a sunset date of August 9, an obituary short enough to fit in its own omnibox.

OpenAI is not throwing Atlas away so much as digesting it: the agentic-browsing work moves into ChatGPT's in-app browser, which picks up multiple tabs, downloads, account logins, and better navigation. The distribution math is what matters. Codex used to be a separate download for people who already knew they wanted a coding agent; now it ships inside the default ChatGPT desktop app on every plan, including Free. That is a very large funnel pointed at agentic coding.

Where a Work session actually runs

The detail most coverage skipped, flagged by Simon Willison, is the split between cloud Work and desktop Work. OpenAI's help center puts it plainly: "Work on web and mobile runs in the cloud. Work in the desktop app can also use local files and desktop apps with your permission. At launch, cloud Work conversations do not appear in desktop Work; desktop Work threads and local files remain on that computer."

Treat them as two colleagues who share a name but not a filing cabinet: the cloud one never touches your local disk, the desktop one can (with permission), and at launch they do not read each other's notes. That split is a real constraint if you are wiring up automations, and a real comfort if your compliance team asks where the files went.

The meter is running

Work is not a seat license. It bills on usage, following the same consumption structure as Codex, with the size of the task, its complexity, and the model you pick driving the cost, and OpenAI has not published per-task rates at launch. Enterprise and Edu admins get spend controls, workspace defaults, group limits, and credit-request workflows in the admin console.

In other words, your chatbot now has a taxi meter. If you are the person who will get asked why the bill doubled, watch week-one consumption closely before you let anyone schedule recurring background tasks, because an agent that happily works for hours will also happily bill for them.

The Cowork collision

None of this is happening in a vacuum. Anthropic pushed Claude Cowork to web and mobile just days earlier, with sessions that keep running server-side after you close the laptop. Two frontier labs shipped the same thesis in the same week: the chat box is dead, long may the delegation box live.

On safety, OpenAI says an Auto-Review layer blocked 100% of attempts to extract protected data during adversarial red-teaming. Take that as a red-team result, not a production guarantee. An agent with connectors into your email, a browser that can operate websites, and permissioned access to local files is a prompt-injection surface of considerable size, and red-teaming numbers have a way of meeting reality once several million users start pasting untrusted content into the loop. Sensible defaults exist (approval gates on sensitive actions, admin-scoped plugin permissions); use them before you point this at anything that can spend money.

Key Takeaways

• ChatGPT Work launched July 9: a GPT-5.6-powered agent that works a goal for hours across connected apps and returns finished sheets, decks, docs, and web apps, with a Sites beta for shareable auto-updating web output.
• OpenAI consolidated its desktop lineup to ship it: the Codex app merges into one unified ChatGPT app (Mac and Windows, all plans including Free), the old app becomes ChatGPT Classic, and the Atlas browser is deprecated on August 9.
• Cloud Work and desktop Work are split at launch: web and mobile sessions run in the cloud with no local file access, desktop sessions can use local files with permission, and the two do not sync.
• Billing is usage-metered like Codex, with no per-task rates published; admins get spend controls, and anyone scheduling recurring agent tasks should watch week-one consumption.
• It lands days after Anthropic's Claude Cowork hit web and mobile, and OpenAI's claim that Auto-Review blocked 100% of data-extraction attempts in red-teaming deserves the usual skepticism: the connector-plus-browser-plus-local-files surface is prompt-injection heaven.

Sources: OpenAI (launch announcement), OpenAI Help Center (cloud vs desktop behavior), The Decoder (connectors, billing, Auto-Review), MacRumors (Codex merger, ChatGPT Classic, rollout), Engadget (desktop app, Atlas deprecation), 9to5Mac (Atlas sunset), Simon Willison (local-files clarification).