What ChatGPT Work really means for your operating model
Today OpenAI shipped ChatGPT Work, an agent built into the ChatGPT subscription your company probably already pays for. It reaches into your connected apps, including Salesforce, Gmail, Google Drive, Slack, and GitHub, and works for hours to hand back a finished spreadsheet, a slide deck, or a working web app. Most of the coverage is about what it can do. The more useful question for anyone running a company is what it changes about how you run one.
OpenAI just handed every employee an autonomous agent. Your operating model decides whether that helps or hurts.
Short answer: ChatGPT Work, released July 9, 2026, is an agent inside ChatGPT that pulls context from your connected apps and runs for hours to produce finished documents, spreadsheets, decks, and web apps. For executives, the release matters because it puts an autonomous agent on every desk before most organizations have built anything to govern one.
We build and run AI-native workflows for operators shipping agents to production, and the gap this release widens is one we hit constantly: teams buy the capability long before they build the controls.
What is ChatGPT Work, and what can it actually do?
ChatGPT Work is an autonomous agent OpenAI has placed inside the ChatGPT app, aimed squarely at the enterprise. It has Codex built in and runs across web, mobile, and desktop, using information from the apps you connect to it.
The scope of what it touches is the part to read twice. At launch, the plugin directory includes Google Drive, SharePoint, Slack, Microsoft Teams, Gmail, Outlook, Salesforce, Zoom, LinkedIn, GitHub, Canva, and Dropbox. It gathers context across those sources and produces finished materials: sheets, slides, documents, reports, and fully built web apps. It stays with a complex project for hours by breaking it into smaller steps, and it can schedule itself to keep working without a person in the chair. Bloomberg described the design as an agent built to field tasks for hours.
It runs on GPT-5.6, released the same day in three sizes: Sol, the flagship; Terra, for everyday work; and Luna, the cheapest. Sol is priced at $5 per million input tokens and $30 per million output tokens, and OpenAI's Sam Altman said it is 54% more token-efficient on agentic coding tasks. The model reached public release today after a delay tied to U.S. government review, which is its own signal about where frontier capability now sits.
Availability tells you how fast this arrives. ChatGPT Work is live today for Pro, Enterprise, and Edu plans, expands to Plus and Business within days, and is reachable free through the Mac and Windows desktop app. This is not a limited beta for a pilot team.
Why does this release matter more than the last agent demo?
The capability is not new. The distribution is.
Autonomous agents have shipped before, from Operator to the first ChatGPT agent. Today a long-running agent lives inside the tool your employees already open every morning, on plans that cover almost everyone with a login. The people who will point it at your Salesforce data sit in finance, sales ops, and marketing, and by next week most of them can turn it on themselves.
This sits on a trend line executives should already know. McKinsey found that the length of task an AI can complete unsupervised has roughly doubled every seven months, reaching around two hours. OpenAI is now advertising "hours" as a headline feature. The demos have been catching up to that number for a while. Today one of them became a default.
It also marks the crossing we have argued for: from an assistant that waits for your prompt to an agent that runs the loop on its own. If you have not made that distinction concrete for your own roadmap, we wrote the operating-model version of it here. The short version is that an assistant speeds up a task a person owns, and an agent owns the task. ChatGPT Work is squarely the second kind.
Was the capability ever your bottleneck?
For two years, AI stalled inside large companies for reasons that rarely came down to the model. The ownership, the access rules, and the review paths were never built.
The readiness data says so plainly. McKinsey's State of AI found 62% of organizations at least experimenting with agents, while only 23% had scaled an agentic system anywhere in the enterprise. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, on escalating cost, unclear business value, and weak risk controls. Every one of those failure modes is an operating-model problem.
ChatGPT Work removes the last comfortable excuse. Leaders could point at the technology and say it was not ready, and often they were right. That line is gone. The harder work remains, the part most companies deferred: deciding who owns an agent's output, what it is allowed to touch, and where a human still signs off. The tool got democratized this morning. The governance around it still has to be built, and that gap is where AI pilots stall.
Mapping this for your own org this week? Download the human-writing skill. It is the free method we use to keep analysis like this from reading like every other AI memo.
What are the three questions ChatGPT Work forces onto your desk?
An agent that can read your CRM and act on its own for hours turns three quiet governance questions into urgent ones. Answer them before the tool answers them for you.
| Question | Why it is urgent now | What good looks like |
|---|---|---|
| Access. What is it allowed to touch? | The default connectors reach Salesforce, Gmail, Drive, and GitHub. Whatever an employee wires in becomes the blast radius of a bad run. | A written policy for which apps can be connected, by whom, for which data. Treat it as data governance, owned like any data policy. |
| Accountability. Who owns what it produces and does? | The agent acts without a person in every step, so the old answer, "the person who ran it," gets thin fast. | A named owner for agent output in each function, decided before the work ships. We covered who holds the mistake here. |
| Oversight. Where is the checkpoint before it acts? | This agent takes actions across your systems, and a checkpoint on the draft never touches them. | A review gate placed before every consequential step in the process. The human-in-the-loop call most execs get wrong. |
The pattern under all three is the same. Guardrails written as polite instructions to the model do not hold when the model is acting on its own, which is why prompt-based guardrails fail. The controls have to live in access, ownership, and review, outside the agent.
What should an executive do in the first 30 days?
The right first move is an access policy, written before shadow usage writes one for you. Assume people on your team are already trying the free desktop version today.
- Inventory the connectors. List which apps employees can wire into ChatGPT Work and set a default for each. Granting an agent read access to Salesforce and Gmail is a data-governance decision made once, for the whole team.
- Name the owners. In the functions most likely to adopt first, finance, sales ops, and marketing, name who answers for agent output before it goes anywhere.
- Instrument one workflow. Pick a single high-value, low-blast-radius workflow, run it with a human checkpoint, and measure it before you widen access. One governed workflow teaches you more than ten ungoverned experiments.
- Set policy for reality. A ban pushes usage into personal accounts you cannot see. A blank check invites the failures in the Gartner data. Aim for the narrow, boring middle: sanctioned access, named owners, and a review gate.
This is ordinary operating-model work that a capable, widely distributed agent finally makes unavoidable.
Frequently asked questions
What is ChatGPT Work? ChatGPT Work is an agent released by OpenAI on July 9, 2026, that lives inside the ChatGPT app. It gathers context from your connected apps and files, then works for hours to produce finished documents, spreadsheets, presentations, and web apps. It is powered by OpenAI's GPT-5.6 model and aimed at enterprise use.
Is ChatGPT Work an agent or an assistant? It is an agent. An assistant responds to a prompt and hands control back. ChatGPT Work plans, acts across your apps, and keeps working toward a goal for hours without a person in every step. That autonomy makes it an operating-model question.
Which apps does ChatGPT Work connect to? At launch, the plugin directory includes Google Drive, SharePoint, Slack, Microsoft Teams, Gmail, Outlook, Salesforce, Zoom, LinkedIn, GitHub, Canva, and Dropbox, among others. The set of apps you connect defines what the agent can read and act on, which is why access policy is the first control to set.
How much does GPT-5.6 cost? GPT-5.6 comes in three sizes, priced per million tokens: Sol, the flagship, is $5 input and $30 output; Terra is $2.50 and $15; Luna is $1 and $6. ChatGPT Work itself is included on Pro, Enterprise, and Edu plans today, with Plus and Business following within days.
What is the main risk for companies adopting ChatGPT Work? The main risk is distribution outrunning governance. A capable agent is now available to nearly every employee, with access to sensitive systems, before most organizations have named owners, set access policy, or built review gates. The failure mode is organizational.
What to decide before your team turns it on
ChatGPT Work is a genuinely capable agent, and by next week most of your employees can run it against your real systems. The tool question is settled. The open question is the one your operating model has to answer: what it can touch, who owns what it does, and where a human still signs off. Decide that this week, because the software already shipped.
We write about what it takes to run an AI-native organization, down to the operating-model calls a leadership team actually has to make. Download the human-writing skill, the free method behind every piece here, or take the lighter door and subscribe for the next one.
Sources
- OpenAI, "ChatGPT is now a partner for your most ambitious work" (July 9, 2026). openai.com
- Bloomberg, "OpenAI Unveils ChatGPT Work Agent to Field Tasks for Hours" (July 9, 2026). bloomberg.com
- Axios, "OpenAI releases GPT-5.6 and ChatGPT Work tool" (July 9, 2026). axios.com
- MacRumors, "OpenAI Debuts ChatGPT Work Agent and New GPT-5.6 Models" (July 9, 2026). macrumors.com
- CNBC, "OpenAI to publicly release GPT-5.6, ending government limits" (July 8, 2026). cnbc.com
- McKinsey, "The State of AI" (2025). mckinsey.com
- McKinsey, "The Agentic Organization: Contours of the Next Paradigm for the AI Era" (2026). mckinsey.com
- Gartner, "Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (June 25, 2025). gartner.com
By Christopher Kliebenstein. We build and run AI-native workflows for operators who are shipping AI to production.