ChatGPT work turns OpenAI's chatbot into a long-horizon agent
OpenAI has repositioned ChatGPT around a new mode, ChatGPT Work, that plans multi-step projects, acts across your connected apps and files, and stays on task for hours before handing back finished documents, decks, and sites. Here's what actually shipped and why the orchestration layer matters more than the model behind it.
OpenAI announced ChatGPT Work on July 9, 2026, and the framing is deliberate: not a smarter chatbot, but an agent that takes a goal, breaks it into steps, and stays with a project for hours until there's a finished document, deck, spreadsheet, or site on the other end.
What shipped
ChatGPT Work runs on GPT‑5.6, OpenAI's new frontier model, and folds in Codex — the coding agent that OpenAI says now has more than 5 million weekly users, over 1 million of whom use it for work outside software development. That last number is the tell: OpenAI is treating Codex's task-decomposition and tool-use machinery as general-purpose infrastructure, not a coding-specific feature, and pointing it at spreadsheets and slide decks instead of pull requests.
The capability set is broad. Plugins connect ChatGPT to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRMs, and project trackers, and a unified directory now surfaces which ones are relevant to a given prompt. Scheduled Tasks let it monitor a channel or dashboard and act on a cadence without you present. On desktop, a built-in browser lets it pull in live web context, and Computer Use lets it click, type, and move files across your local apps in the background. A new public beta, Sites, lets ChatGPT turn a workflow into a shareable interactive web app — a live dashboard or launch tracker, not just a static export. Notably, the standalone Atlas browser is being sunset in favor of folding that capability into the desktop app, and the Codex app itself is merging into it, with inline diff editing and PR review now living alongside chat and work modes.
Rollout is Pro, Enterprise, and Edu plans today, with Plus and Business following in the next few days. The desktop app's Chat, Work, and Codex modes are available on every plan, including Free.
The part worth studying: the orchestration layer
What's actually new here isn't the model — it's the scaffolding around it. Every example OpenAI cites follows the same shape: a goal comes in, gets decomposed into steps, pulls context from connected systems, takes action through a browser or the OS, pauses at a checkpoint for approval, and emits an artifact.
The case studies in the announcement are useful precisely because they're mundane: Zapier used it to trace CRM and email touchpoints into a weekly executive dashboard that surfaced seven figures in missed pipeline. RingCentral turned manual monthly launch reviews into a workflow that lets one R&D efficiency manager, Vaneet Seth, support roughly 50 product managers instead of one. Virgin Atlantic compressed weeks of competitive customer-experience research into hours. NVIDIA cut GTC prep time by automating the account and meeting tracking that used to eat 40% of one manager's pre-event schedule. None of these are novel LLM capabilities — they're integration and persistence problems, solved by wiring a planner to a connector graph with a human checkpoint in the loop.
Where this sits competitively
This is the same architecture Microsoft is building into Copilot agents and that Anthropic ships through Claude's computer use and Claude Code — a planner, a tool graph, durable state across a long task, and an approval gate before consequential actions. The frontier-model layer is converging; the differentiation is shifting to how well a vendor's connector ecosystem, checkpoint design, and audit trail hold up when an agent is running unsupervised for hours inside a real company's systems. That's a harder problem than benchmark accuracy, and it's the one to watch as these products move from demos to production.