OpenAI has introduced ChatGPT Work, an agentic workspace inside ChatGPT intended to take on extended, multi-step business projects rather than simply answer individual prompts. Announced July 9, the product can gather context from connected apps and files, create documents and other outputs, take sequences of actions, ask for approval when needed, and continue work over hours. [1]
The launch is significant because it pushes ChatGPT further into competition with workplace software makers and agent platforms, including Microsoft 365 Copilot Cowork and Anthropic Claude Cowork. OpenAI is betting that its consumer reach, Codex-derived agent technology and connector ecosystem can turn ChatGPT from a conversational assistant into a general-purpose execution layer for knowledge work.
From prompts to long-running projects
ChatGPT Work is powered by GPT-5.6 and incorporates technology from Codex, OpenAI’s agentic coding product. The central distinction is persistence: users can describe an objective, provide constraints and let Work break the project into smaller steps, report its progress, pause for a decision, and resume. [1]
OpenAI says Work can produce reports, spreadsheets, presentations and documents, while drawing on files, websites and connected workplace systems. Its new Sites feature, initially in public beta, can turn an idea or work product into an interactive website or web application that is shareable by URL. Suggested applications include internal portals, project trackers, dashboards, launch calendars and interactive reports. [1]
Scheduled Tasks extends the agent model beyond a single session. A task can run once, recur on a timetable, activate on an event, or monitor a source for changes. OpenAI’s examples include reviewing Slack activity, watching customer feedback, checking a dashboard or updating a presentation when new email arrives. The design moves closer to delegated workflow automation, where the AI is expected to act across time instead of waiting for another prompt. [1]
OpenAI says more than 5 million people use Codex weekly, including more than 1 million people using it for work outside software development. That claim helps explain why the company is positioning Work as an expansion of agentic capabilities first built for coding into finance, sales, operations, research and other knowledge-work functions. [1]
Connectors, browser access and the desktop app
Work can use OpenAI’s unified plugins directory, which the company says contains more than 1,400 plugins. Listed integrations include Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, Salesforce and project-tracking systems. A user can explicitly call a connector with an “@” mention, or ChatGPT can select an applicable plugin automatically. [1]
On desktop, authorized local files and applications can become part of the agent’s working context. OpenAI says local files and outputs remain on the computer unless a user explicitly moves or shares them, while cloud Work sessions synchronize across web, mobile and desktop. [2]
The desktop application also includes a browser intended to let Work gather information from sites, interact with web tools and work with online files without forcing users to move among separate applications. OpenAI is also updating its Chrome extension to surface ChatGPT in Chrome’s sidebar. [1]
For organizations, the attraction is straightforward: an agent with access to the documents, messages, calendars and systems where operational context already resides could reduce the manual effort of assembling information before a task even begins. It also creates the central enterprise challenge: the same access that makes an agent useful can expose sensitive data or allow consequential actions if permissions and review policies are poorly configured.
Availability, plans and product reorganization
Work began rolling out July 9 on web and mobile for Pro, Enterprise and Edu users. OpenAI said Plus and Business users would receive access over the following days. The refreshed Mac and Windows desktop app is available globally, with Chat, Work and Codex offered across plans, including Free, subject to feature and workspace restrictions. [1]
OpenAI has reorganized its desktop products around the new application. The ChatGPT desktop app now combines Chat, Work and Codex; the former standalone Codex app is being merged into it; and the previous ChatGPT desktop app has been renamed ChatGPT Classic. Codex remains the more technical, developer-oriented environment, while Work is aimed at broader workplace tasks. [1]
Usage follows the more variable structure used for Codex rather than a simple count of ordinary chat exchanges. OpenAI warns that longer or more complex jobs can consume substantially more included capacity. Enterprise and Edu administrators can establish controls at workspace, group and individual levels. [1]
That pricing and capacity model may be difficult for ordinary users to predict. Axios noted that GPT-5.6 includes Sol, Terra and Luna configurations with differing speed, capability and cost profiles. Developer Simon Willison, cited by Axios, estimated that the same prompt could cost roughly 0.71 cents to 48.55 cents depending on model and reasoning selections. [5] The figures illustrate an important practical issue for agents: the cost of a multi-step workflow depends not only on the user’s request, but also on how long the system reasons, browses, calls tools and retries work.
Governance is central to the enterprise pitch
OpenAI has built Work around the premise that users should be able to monitor and intervene in delegated tasks. Users can review progress, redirect the agent and choose when approval is required. Enterprise and Edu administrators can govern access to connected tools, company context, browser and network capabilities, and sensitive actions. The company’s Compliance API is intended to provide visibility into Work conversations and actions. [1]
OpenAI also describes an auto-review layer that uses advanced models to inspect important actions involving connected tools and APIs before they are carried out. These controls are designed to reduce risks such as an agent sending information to the wrong recipient, acting on incomplete context or making an unauthorized change in a connected system. [1]
However, these safeguards do not answer every operational question. A polished report can still contain incorrect analysis; a workflow can follow approved permissions while making a poor business judgment; and broad connector access can make an accidental disclosure more consequential. OpenAI has not published independent failure-rate, accuracy or productivity benchmarks for Work. Organizations evaluating the product will need to decide which tasks can run autonomously and which require human review before outputs are shared or actions are executed.
Early productivity claims and a crowded agent market
OpenAI says nearly all of its internal teams, including finance and sales, use ChatGPT Work and Codex. It cites a sales process that turned a discovery conversation into a tailored proof of concept within 24 hours instead of weeks, and finance workflows that reduced month-end close and forecasting tasks from days to hours. [1]
The company also points to early users at NVIDIA, Shopify, RingCentral and Zapier. OpenAI says NVIDIA’s Will Daney automated conference account, registration and feedback analysis; Shopify’s Chris Jones used Work to analyze AI adoption among 3,500 non-R&D employees; and Zapier’s Angela Ferrante used it to analyze thousands of leads and identify seven figures in potential sales. [1] These are vendor-reported examples rather than independently audited studies, so they indicate possible use cases more than proven, broadly replicable savings.
OpenAI’s own research on Codex similarly points to adoption rather than verified productivity outcomes. By May, 80.6% of sampled individual users had made at least one request estimated to represent more than 30 minutes of human work, while 25.6% had made a request estimated to exceed eight hours. Those measures describe estimated task duration and usage patterns, not confirmed labor hours saved. [3]
The competitive stakes are high. Anthropic has positioned Claude Cowork around agentic work for non-coders, while Microsoft’s advantage lies in its deep integration with Microsoft 365 data, documents and enterprise processes. Google also has a natural distribution channel through its productivity suite and workplace AI initiatives. Industry analysts have increasingly characterized this category as a shift from prompting toward persistent agents that can plan, use tools and pursue an outcome across many steps. [4]
A powerful launch with adoption friction
Work gives OpenAI a more complete answer to the question of what ChatGPT should do after generating text: it should assemble context, build a deliverable, execute routine steps and remain available as a collaborator over the life of a project. If the agent proves dependable, that could alter the role of many point tools and lightweight workflow products.
Yet the company’s simultaneous desktop redesign has drawn criticism. TechRadar reported complaints over the merger of ChatGPT and Codex, the ChatGPT Classic rebrand, revised navigation and difficulty locating prior conversations, projects and custom GPTs. Apple commentator John Gruber described the redesign as a major failure and noted that the Mac app had grown from roughly 159 MB to 1.5 GB. [6]
The contrast captures the challenge facing OpenAI. ChatGPT Work offers a broad and ambitious vision of AI-mediated workplace execution, but its value depends on more than model capability. Users will need intelligible interfaces, predictable costs, reliable permissions, transparent review trails and confidence that the agent’s outputs are correct. The launch establishes OpenAI as a serious contender in that race; sustained adoption will depend on whether the product can make complex delegation feel more dependable than complex.
Editor’s Take
I see ChatGPT Work as a meaningful product shift, not simply another chatbot feature. The useful promise is the ability to hand an agent a bounded business objective—assemble research, reconcile inputs, draft a presentation, monitor a source—and stay in the review loop rather than manually stitching together dozens of small prompts. If the connectors and approval controls work reliably, this could remove a lot of the tedious context-gathering that slows finance, sales, operations and product teams.
The practical test is not whether Work can generate an impressive report, but whether it can be trusted with messy company data and repeatable workflows. Enterprises should begin with read-heavy, reversible jobs: weekly summaries, lead research, dashboard monitoring and draft creation. They should be far more cautious about sending emails, changing records or making decisions from incomplete evidence. OpenAI’s customer examples are promising, but they are still vendor-reported anecdotes; the next thing to watch is independent evidence on error rates, review burden, connector security and the real cost of long-running tasks.
References
- OpenAI – https://openai.com/index/chatgpt-for-your-most-ambitious-work/
- OpenAI Help Center – https://help.openai.com/en/articles/20001275/
- OpenAI, “How agents are transforming work” – https://openai.com/index/how-agents-are-transforming-work/
- Forrester, “Persistent AI Agents Will Rewire How Enterprises Execute Work” – https://www.forrester.com/report/persistent-ai-agents-will-rewire-how-enterprises-execute-work/RES196634
- Axios – https://www.axios.com/2026/07/12/openai-chatgpt-work-luna-terra-sol
- TechRadar – https://www.techradar.com/ai-platforms-assistants/christ-almighty-this-is-so-bad-chatgpts-big-app-update-brings-huge-changes-to-your-workflows-and-users-seem-to-hate-it
,_HomeStreet_Bank_Headquarters_--_2022_--_1454.jpg){kind=link}
