OpenAI released ChatGPT Work in July 2026. It's a mode inside ChatGPT that runs long, multi-step tasks on its own, reaching into connected apps, local files, and a browser instead of answering one question at a time. It runs on GPT-5.6, and it ships with three capabilities that didn't exist in the chat interface: Scheduled Tasks, Computer Use, and Sites.
The practical difference is duration. This is built for work that takes hours. That changes how your plan's usage gets consumed, and for Enterprise and Edu organizations it changes what admins need to configure before rolling it out.
What changed about the interaction model
ChatGPT has worked in turns: you send a message, it responds, the exchange ends. That handles drafting an email or summarizing a document. It handles poorly anything that needs steps in sequence, a pause while an external input arrives, a loop back when something changes, or a process that runs while you're doing something else.
ChatGPT Work covers that gap. OpenAI describes it as built to "take on more ambitious tasks" — gathering information across several apps, producing finished materials like slides, sheets, docs, and web apps, and staying with a project for hours by breaking it into smaller pieces and working through them without supervision.
The word doing the work in that description is "without supervision." The system holds a multi-step plan, uses tools to execute parts of it, checks its own output, and returns a result at the end.
Where it came from
Codex was OpenAI's coding agent, built for developers to write, debug, and ship code with less manual intervention. Its architecture assumed multi-step tasks, tool use, and long time horizons from the start.
OpenAI has folded Codex into the main ChatGPT desktop app. That app now contains three modes — Chat, Work, and Codex — in one application. The separate Codex app is becoming the new ChatGPT desktop app; existing Codex users update as normal, and desktop Codex projects stay reachable from the ChatGPT mobile app.
ChatGPT Work inherits that architecture, which is what lets it run long. OpenAI reports more than 5 million weekly Codex users, with more than 1 million now using it for work outside software development. The audience for agentic tooling had already grown past developers before this release packaged it for them.
What it can do
Plugins are the connection layer. ChatGPT Work links to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRMs, and project trackers. You can point it at a specific source by typing "@" and the app name, or leave it to pick relevant sources itself.
From those sources it produces slides, sheets, docs, analyses, and web content. The central use case is converting raw input — discovery call notes, source data, customer research — into material a team can use or review.
Computer Use is desktop-only. The agent operates your machine directly: clicking, typing, moving between apps, handling files. OpenAI's framing is that it works in the background across your apps, tools, and browser. That covers local software with no plugin and no API.
Scheduled Tasks is the piece that separates this from a chat window. Work can run on a schedule or fire on an event rather than waiting for a prompt. OpenAI's examples: reviewing the week's Slack activity and refreshing a recurring meeting agenda, checking sites and dashboards each morning and reporting what moved, watching incoming customer feedback and turning repeated themes into prioritized product ideas. You configure it once and it runs without you.
Sites, in public beta, turns work into an interactive page or web app shared by URL — live dashboards, project trackers, launch calendars, prototypes, interactive reports. OpenAI says ChatGPT can keep a Site current as the underlying information changes.
The desktop app also carries a built-in browser for researching markets, pulling from websites, comparing sources, and opening Google Workspace and Microsoft 365 files in place.
The model underneath
GPT-5.6 arrived alongside this release in three tiers: Sol as the flagship, Terra as the balanced everyday model, and Luna as the cost-efficient option. Which one you get depends on your plan:
- Free and Go: GPT-5.6 Terra
- Plus, Pro, Business, and Enterprise: a choice of Sol, Terra, or Luna, each with a settable effort level
max, a higher-reasoning setting, available to anyone with GPT-5.6 access in Work and Codex, toggled on in settingsultra, the highest-capability setting, which coordinates several agents across parallel workstreams, available in Work to Pro and Enterprise
The model tier matters more here than it does in chat. Agentic work needs a model that can reason through intermediate steps without losing the original goal, absorb ambiguous instructions, and produce output matching a supplied template. OpenAI positions GPT-5.6 as built for exactly that combination.
How usage works
Read the usage model before rolling this out to a team. OpenAI states in the announcement that Work is meant for longer and more involved jobs than an ordinary chat request, so usage is counted differently: the amount consumed scales with the amount of work performed, and heavier tasks draw more of the allowance included in your plan.
In practice, your plan includes an allowance. A quick message uses a sliver of it. A Work task that runs for hours, makes dozens of tool calls, and assembles a multi-part output uses proportionally more. The allowance itself doesn't shrink; the rate at which a complex task draws it down is what differs.
There's no new pricing tier and OpenAI announced no price increase with this release. It's an accounting consequence of agentic work being computationally heavier than a single-turn conversation.
For an individual on Plus or Pro this is unlikely to register day to day unless you're running large batches of long tasks. For a team deploying at scale it belongs in the adoption plan.
Admin controls for Enterprise and Edu
Enterprise and Edu admins get dedicated tools for managing Work usage as it spreads across teams. The spend controls live in the Admin Console and cover workspace-level defaults, group-level limits, individual overrides for people who need more capacity, and a review queue for additional-credit requests that carries the requester's project details and rationale.
This is admin-level only. Individual Plus and Pro plans have no Admin Console and no organizational usage management.
OpenAI's stated intent is that admins can back high-impact work across teams without raising limits for everyone. That's governance rather than cost-cutting — the point is directing capacity toward work that justifies it, instead of choosing between restricting everybody and opening the taps.
There's also Auto-review, where a separate model inspects significant actions involving connected tools and APIs before they execute, to prevent sensitive information being shared without authorization. For any organization running client data or proprietary material through connected CRMs and document systems, that's the control that matters most.
Who gets what
Coverage of this release has been muddled on availability, so here is the breakdown as OpenAI stated it at launch.
The desktop app for macOS and Windows carries Chat, Work, and Codex on every plan including Free, available globally. On web and mobile, Work went first to Pro, Enterprise, and Edu, with Plus and Business following within days. Sites is in public beta. Computer Use is desktop-only, inside the desktop app that all plans can download. Admin spend controls are limited to Enterprise and Edu admins through the Admin Console.
The desktop distinction is the useful one: a Free-plan user who downloads the desktop app has ChatGPT Work today, running GPT-5.6 Terra.
What to do with it
The agent framing here is accurate. Work runs for hours without human input, uses tools, and returns finished output, which is a real departure from what ChatGPT was.
Its value scales with what you connect it to. An unplugged Work session is capable and isolated. Wired into your CRM, project tracker, Slack, and Drive, it becomes something that can synthesize across systems. The plugin setup isn't an optional extra step; it's where the capability comes from.
Scheduled Tasks is probably the highest-leverage feature for most teams. One-off tasks help. Recurring ones that refresh on their own — briefings, monitored dashboards, weekly summaries — compound. If your team has recurring information workflows, start there.
Sites is early. Public beta means the feature set will move and stability isn't guaranteed, so it's worth watching and too soon to build a dependency on.
Enterprise admins need a deployment plan before a broad rollout. Because complex tasks draw down allowances proportionally, an ungoverned rollout produces uneven consumption — some teams lean on it hard, others never touch it. The spend controls address that, but only if someone configures them.
The open question
Structurally, the interesting part is distribution. Codex was always the underlying technology; putting it in the main product, on every plan, on desktop, without a separate download, drops the adoption barrier considerably. OpenAI isn't positioning Work as a developer tool. The use cases in the announcement are sales teams building proof-of-concept materials, finance teams accelerating month-end close, marketing teams turning research into campaign assets. None of those are technical users.
Whether it holds up at that scale — whether Work performs as reliably inside a real organization as it does in a controlled demo — is what the next several months of adoption will answer. The capabilities are real and shipped. Reliability in production is still being established.
If your team has been treating ChatGPT as a chat interface, the thing worth reconsidering is which recurring, multi-step work you still do by hand because no tool could hold the whole context and carry out the steps. That's the shape of problem this was built for.
