ChatGPT Work: OpenAI’s AI Agent That Finishes the Job
ChatGPT Work is OpenAI’s autonomous AI agent, launched July 9, 2026, that goes beyond answering questions — it delivers finished documents, spreadsheets, presentations, and websites. It is the defining product of the chat gpt 5.6 era, powered by GPT-5.6, and marks a shift from chatbot to AI coworker. See the official ChatGPT Work product page. Unlike regular ChatGPT, the Work agent takes a high-level goal, builds a step-by-step plan for your approval, then executes across your connected apps and files — staying with complex projects for hours. The output is finished material, not a chat response.
What Is ChatGPT Work?
Launched alongside GPT-5.6, ChatGPT Work is a dedicated mode within ChatGPT for longer, multi-step projects. OpenAI built it as a direct evolution of Codex — the standalone coding agent is now embedded in the Work agent rather than a separate product. OpenAI reports 5 million+ weekly Codex users at launch, more than 1 million of them for non-development tasks. Activated via a “Work” toggle, it produces finished files — spreadsheets, slide decks, reports, full web apps — by autonomously acting across connected tools, rather than waiting for your next message. Asking regular ChatGPT to “compare our top 5 competitors” returns a text summary; asking ChatGPT Work returns a structured spreadsheet populated with live data, ready to present.
“ChatGPT Work is designed to do the kind of work that used to require a team — research, analysis, coding, building — and deliver it finished, not as a draft.”
OpenAI — “ChatGPT is now a partner for your most ambitious work,” July 9, 2026
Key Features
| Feature | What It Does | Key Detail |
|---|---|---|
| Plan Mode | Shows step-by-step plan before any action | Requires explicit approval |
| Scheduled Tasks | Runs on schedules or triggered by changes | Replaces manual recurring work |
| Computer Use | Browses web, fills forms, acts on desktop | Works without APIs |
| Sites | Builds and deploys interactive web apps | No code required |
| Plugin Directory | 1,400+ integrations | Covers most enterprise tools |
Plan Mode generates a reviewable plan before the agent touches anything — you can edit, remove, or reorder steps, then approve, eliminating the unsupervised-action problem of earlier agents. Scheduled Tasks let Work run once, repeat on a schedule, or trigger when monitored conditions change; OpenAI’s finance team used them to compress month-end close from days to hours. Computer Use lets the agent browse, fill forms, and interact with desktop apps that have no API, extended by 1,400+ plugins. Sites — launched in public beta alongside Work — turns a prompt into a live, shareable web app with no coding.
Access and Pricing
ChatGPT Work is exclusive to paid plans; Free and Go are excluded. Pro, Pro Lite, Enterprise, and Edu received access on day one (July 9, 2026); Plus and Business followed within days. The macOS desktop app is free to download, but the Work feature inside requires a paid subscription; Windows rolled out within days. Pricing is usage-metered, drawing from a credit pool following the Codex credit structure rather than flat-rate chat pricing — heavier multi-hour tasks consume more credits. Exact per-task rates have not been publicly disclosed.
Real-World Results
- Zapier reduced lead inspection from 35–45 minutes per lead to near-automated, surfacing seven figures per month in missed pipeline.
- NVIDIA GTM Manager Will Daney automated 40% of his pre-event spreadsheet prep for the GTC conference.
- Virgin Atlantic cut competitive research from weeks to hours.
- RingCentral scaled one product manager from ~6 customers to ~50.
Nearly all of OpenAI’s own internal teams use ChatGPT Work for part of their workflows, and its sales team produced a full proof-of-concept in 24 hours — previously a multi-week task. The Work agent integrates natively with Zapier, opening access to thousands of additional automation workflows without custom API work.
ChatGPT Work vs. Claude Cowork and Copilot Cowork
Anthropic’s Claude Cowork reached the market first — its launch triggered a measurable selloff in software and professional-services stocks. Microsoft’s Copilot Cowork followed. ChatGPT Work entered July 9, 2026, as OpenAI’s answer, backed by GPT-5.6 and 1 billion+ existing ChatGPT accounts. All three target the same workflow — a complex multi-step task in, a finished output out — and differ on integration depth, model capability, and output types. Work leads on breadth (code, Sites, documents, spreadsheets in one product), Codex integration, 1,400+ plugins, and the explicit Plan Mode approval flow. Claude Cowork is noted for strong reasoning and document handling; Copilot Cowork benefits from deep Microsoft 365 integration. The choice usually comes down to your existing toolchain. Within Work, the OpenAI Luna model handles high-volume, fast-turnaround tasks while Sol powers the most demanding sustained reasoning.
Enterprise Security and Governance
OpenAI reports its Auto-review safeguard blocked 100% of protected-data extraction attempts in internal red-team testing, gating planned actions before system access so the agent cannot exfiltrate connected data without triggering review. Enterprise and Edu customers got a 2-week opt-out preview before activation. The Compliance API provides audit trails and policy controls for regulated industries — logging every action, enforcing data boundaries, and reviewing execution history. This pairs with how programmatic tool calling extends governance to API-level integrations for teams building on top of Work rather than using the consumer interface.
How to Get Started
- Confirm your plan — Pro, Pro Lite, Enterprise, Edu, Plus, or Business.
- Open ChatGPT on web or desktop and switch to Work mode via the toggle (top of chat on web/mobile; ChatGPT menu on macOS).
- Write an outcome-defined prompt — e.g. “Build a competitive analysis spreadsheet comparing our top 5 rivals on pricing, features, and market share.” Vague prompts produce vague plans.
- Review the Plan Mode steps, editing or reordering anything off-intent, then approve and monitor.
- Collect and verify the output — for critical work, spot-check the data sources the agent referenced.