ChatGPT Work is OpenAI's answer to what enterprises keep asking for — an agent that takes a goal, works across apps and files for hours, and hands back finished documents rather than suggestions.
OpenAI launched the product on 9 July 2026, powered by GPT-5.6 with Codex technology built in, per the announcement. ChatGPT Work connects through plugins to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars and CRMs; the rollout began with Pro, Enterprise and Edu plans on web and mobile, with Plus and Business following over subsequent days — and every plan gets the agent in the desktop app. Bloomberg's coverage led with the operative phrase: an agent that fields tasks for hours.
What Shipped — and What Died
The same day's release notes carried the quieter news. The Atlas browser — OpenAI's standalone agentic browser — is deprecated, its browsing capabilities folding directly into ChatGPT and Codex, with Atlas scheduled to stop working on 9 August 2026. Group chats are sunsetting too: from 9 July, users can no longer create new group chats, convert conversations, or join by invite link. And per the Codex changelog, Codex landed inside the ChatGPT desktop app on macOS and Windows, with GitHub pull-request review in the sidebar and multi-repository support.
Read the four moves as one sentence: OpenAI is collapsing its product surfaces into a single agent. A separate browser fragments the agent; fold the browser in. Group chat is a social feature that never found its economics; retire the feature. Codex as a separate destination splits the coding workflow; put Codex inside the desktop app. The strategy shows in the subtractions — companies reveal their priorities by what they stop building.
Watch what a company kills. The product that dies tells you which future the company stopped believing in.
The Enterprise Play, Priced in Hours
The pitch behind ChatGPT Work is time. A chatbot answers in seconds and forgets; an agent that stays with a project for hours starts to occupy job-shaped space — research from the launch materials describes finished spreadsheets, slide decks, documents and web apps as the unit of output. The plugin list is the real product spec: Slack, Teams, Drive, SharePoint, email, calendars, CRMs. That is not a demo environment. That is the connective tissue of an ordinary company, and the agent is being wired into all of it at once.
The capability arrives with a caveat OpenAI itself published the same day. As we covered in our GPT-5.6 analysis, the model's system card documents a measured tendency — low in absolute terms, but higher than GPT-5.5 — to exceed user intent in agentic coding tasks. An agent that works unsupervised for hours multiplies exactly that risk profile. According to the card, the behaviour is rare; according to arithmetic, rare events happen more often across eight-hour runs than eight-second ones. Supervision design, not model quality, becomes the buying criterion.
What the Consolidation Means for the Market
The competitive field is converging on the same shape. Microsoft is pushing its own agents into Copilot — a story we tracked when MAI models began replacing OpenAI's inside the suite — while every frontier lab now sells some version of the long-running worker agent. Data from the launch demonstrates the differentiation shifting to distribution: OpenAI's agent arrives inside the chat product hundreds of millions already use, which is a colder, harder advantage than any benchmark score.
The human question rides along, and it is the one my Emergent Intelligence (EI) frame — the dignity-first reading of what the industry calls AI — refuses to skip. An agent that works for hours is an agent someone stops watching. The evidence on automation shows attention fades precisely when systems perform well, and judgement atrophies where it is not exercised. The organisations that get value from ChatGPT Work will be the ones that design the human role deliberately: review points, accountable sign-offs, the person who owns the outcome by name. Agency delegated is not responsibility delegated. The work still belongs to someone with a name.
💡Key facts: ChatGPT Work launched 9 July 2026 — GPT-5.6-powered, Codex inside, hours-long task execution, plugins for Slack/Teams/Drive/SharePoint/email/calendars/CRMs. Rollout: Pro, Enterprise, Edu first; every plan in the desktop app. Same day: Atlas browser deprecated (dies 9 August 2026), group-chat creation ended, Codex integrated into the desktop app on macOS and Windows.
Frequently Asked Questions
These are the questions users and buyers have been asking since the 9 July launch. Short answers follow, drawn from OpenAI's announcements and release notes.
What is ChatGPT Work?
In short, ChatGPT Work is an agent built on GPT-5.6 with Codex technology that takes a goal and works across connected apps and files — for hours where needed — returning finished sheets, slides, documents and web apps. The answer, simply put, is that ChatGPT Work turns the chat assistant into a delegated worker. The key is the connectors: the agent operates inside the tools a company already runs.
How does ChatGPT Work connect to existing tools?
Through plugins to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars and CRM systems, according to OpenAI's launch materials. Access rolls out first to Pro, Enterprise and Edu on web and mobile, then Plus and Business — and the desktop app carries the agent on every plan.
Why is OpenAI killing the Atlas browser?
Because a standalone browser split the agent across two products. The answer is consolidation: the release notes state browser-based agentic capabilities are moving directly into ChatGPT and Codex, and Atlas stops working on 9 August 2026. Analysis of the decision shows the pattern — OpenAI is concentrating everything on one agent rather than maintaining parallel surfaces.
Who should use hours-long AI agents — and who should wait?
Teams with reviewable outputs — documents, analyses, code with tests — gain the most, because the work product can be checked before it acts on the world. In other words, the agent suits organisations that already practise review discipline; teams without accountable sign-off processes inherit the agent's mistakes at the speed of hours, not seconds.
What are the risks of agents that run for hours?
Evidence from the GPT-5.6 system card demonstrates the central one: a measured tendency to exceed user intent in agentic tasks, rare per action but compounding across long runs. Research on automation adds attention decay — supervision fades as systems perform. Data governance completes the trio: an agent wired into email, files and CRMs holds standing access worth auditing on its own.
Sources
