ChatGPT Work: OpenAI’s New AI Agent Transforms the Workplace
The $850 billion question isn’t whether AI can do your job. It’s whether enterprises will actually pay for it at scale and on July 9, 2026, OpenAI bet its IPO on the answer being yes.
That’s the day the company launched ChatGPT Work, a new mode built into ChatGPT that doesn’t just answer questions but goes off and does things: reading your Slack channels, checking your calendar, drafting your emails, reviewing pull requests, and building finished documents while you’re at lunch. It arrived alongside the GPT-5.6 model family and a quiet but significant merger Codex, OpenAI’s coding agent, is now folded directly into the ChatGPT desktop app.
From Chatbot to Autonomous Worker
For most of its life, ChatGPT has been reactive: you ask, it answers. ChatGPT Work flips that relationship. Instead of producing a single response, it plans a sequence of steps, executes them across the apps you’ve connected, and checks back in with a finished artifact a spreadsheet, a slide deck, a scheduled meeting, a merged pull request.
The Persistent Cloud VM Architecture
One of the more technical, and more consequential, details to come out of launch coverage is that ChatGPT Work runs on a persistent cloud virtual machine rather than executing locally on your device. That means a task you kick off on your laptop keeps running in the cloud even after you close the lid, and you can check on it — or start something new — from your phone.
This is a real architectural fork from competitors. Claude Cowork, by contrast, runs on your own machine, which has upsides for data locality but means the agent generally isn’t working while your computer is asleep. OpenAI’s cloud-first bet trades some of that local control for “always-on” availability across devices — you can, per OpenAI’s own framing, start building a website from your phone at the beach and pick it back up on your desktop later.
The Three Modes: Chat, Work, and Codex
The redesigned ChatGPT desktop app now has a visible toggle between three modes:
- Chat the familiar conversational ChatGPT experience
- Work the new agentic mode covered in this guide
- Codex OpenAI’s coding agent, now unified into the same app rather than living as a separate product
Andrew Ambrosino, who leads Codex at OpenAI, described the merger as “only the first” step toward unifying the app experience across web, mobile, and desktop suggesting more consolidation is coming rather than less.
How ChatGPT Work Works
MCP Plugins: The Integration Layer
ChatGPT Work connects to your existing tools through the Model Context Protocol (MCP) an open standard for letting an AI model talk to external software. Think of MCP as something like USB-C for AI agents: instead of OpenAI building a custom, one-off connector for every single app, MCP gives developers a common plug that any compliant tool can use.
That matters for two reasons. First, it means the integration list can grow faster, since third parties can build MCP servers for their own products rather than waiting on OpenAI. Second, it signals that MCP originally an Anthropic-authored standard is becoming a shared, cross-vendor convention rather than something tied to one company’s ecosystem.
At launch, ChatGPT Work connects to more than a dozen apps, including:
| App | What ChatGPT Work Can Do |
|---|---|
| Gmail | Read, reference, and draft replies |
| Google Calendar | Create events, manage scheduling, and flag conflicts |
| Slack | Read channels, summarize conversations, and draft messages |
| GitHub | Triage pull requests and review code |
| Google Drive | Access and reference documents, spreadsheets, and presentations |
| Outlook | Triage inboxes and draft replies |
| Microsoft Teams | Summarize conversations and prepare follow-ups |
| SharePoint | Summarize sites and access files |
| Salesforce | Query and update records |
| Adobe | Assist with design workflows and editing tasks |
| Zoom | Surface meeting insights and summaries |
| Support professional networking tasks | |
| Canva | Search, create, and edit designs |
| Dropbox | Access, save, and share files |
The GPT-5.6 Brain Behind ChatGPT Work
The foundation of ChatGPT Work is GPT-5.6, OpenAI’s latest model family designed to power advanced AI agents.
The model family reportedly includes three variants, internally referred to as Sol, Luna, and Terra. Each version is optimized for different needs, including maximum capability, faster responses, and balanced everyday performance.
However, for most professionals, the important difference is not the model name. The major improvement is the ability to maintain context across extended tasks.
Traditional AI conversations often lose track of previous details after multiple exchanges. ChatGPT Work is designed to follow ongoing projects over hours or even days, allowing users to continue complex workflows without rebuilding the background information every time.
For knowledge workers, this means AI can become less like a search assistant and more like a long-term project partner.
Scheduled Tasks and Autonomous Execution
One of the biggest changes in ChatGPT Work is its ability to continue working after the user leaves the application.
Through scheduled workflows, the AI agent can perform recurring tasks such as:
- monitoring Slack or Teams conversations
- collecting relevant updates
- organizing information into documents
- notifying users when important changes occur
For example, instead of manually checking multiple channels every morning, a team member could set up an AI workflow that gathers important updates and prepares a summary automatically.
This type of background assistance moves AI from a reactive tool into a more proactive workplace system.
ChatGPT Work Features Explained
Document and Asset Creation
ChatGPT Work can create professional assets beyond simple answers and summaries, including spreadsheets, presentations, reports, research documents, and shareable websites from prompts.
Role-Based Onboarding
The platform adapts workflows based on user roles. Product managers can focus on research and project management, while sales teams can manage customer tracking and communication tasks. Users can also organize work into Projects to keep related files, chats, and tasks together.
Mobile-First Experience
ChatGPT Work is designed for flexible productivity, allowing users to start tasks, review progress, and manage workflows directly from mobile devices.
ChatGPT Work Pricing: Every Plan Explained
Here’s the tier-by-tier breakdown, based on OpenAI’s published plans:
| Plan | Price | ChatGPT Work Access | Best For |
|---|---|---|---|
| Free | $0 | Desktop app only (Chat/Work/Codex toggle) | Trying it out |
| Plus | $20/month | Full web, mobile, and desktop access | Individual professionals |
| Pro | $100–$200/month | Full access, 5x–20x Plus usage limits | Heavy and power users |
| Business | $20–$30/user/month | Full access, team tools, 60+ connectors | Teams of 2+ |
| Enterprise | Custom pricing | Full access, unlimited usage, SCIM, SLA | Large organizations |
Hidden Costs to Consider
The advertised subscription price isn’t necessarily the full bill. Watch for:
- Codex usage billing above the baseline included in your plan
- Separate API costs if you’re building on top of ChatGPT Work rather than just using it
- Custom MCP integration development, which can run anywhere from roughly $10,000 to $50,000+ for a bespoke internal connector
- Trade-offs between enterprise knowledge management settings and connector access, which IT teams will want to map out before a broad rollout
ChatGPT Work vs. The Competition
OpenAI isn’t launching into a vacuum. Anthropic’s Claude Cowork went generally available in April 2026, and Microsoft’s Copilot Cowork launched in mid-June 2026 — meaning ChatGPT Work is the third major “AI coworker” to ship inside of about ninety days.
Three-Way Feature Comparison
| Feature | ChatGPT Work | Claude Cowork | Microsoft Copilot Cowork |
|---|---|---|---|
| Entry price | $20/month (Plus) | Requires Pro/Max subscription | Requires Microsoft 365 E5/E7 |
| Architecture | Persistent cloud VM | Runs locally on your machine | Microsoft cloud |
| Mobile support | Full | Limited | Limited |
| Integrations | 14+ via MCP at launch | Broad third-party support | Deep, native Microsoft 365 |
| Underlying model | GPT-5.6 | Claude | Claude (via Anthropic partnership) |
| Best fit | Cross-platform teams already spread across many tools | Individual power users who want local control | Organizations already deep in Microsoft 365 |
Is ChatGPT Work Safe? Enterprise Security Deep-Dive
Data Privacy Controls
OpenAI says Enterprise accounts get Zero Data Retention (ZDR), and users across tiers can opt out of having their data used for model training. That’s a meaningful baseline, but it’s worth being precise about what it does and doesn’t cover ZDR governs how OpenAI itself stores and uses data, not how each connected third-party app (Slack, Salesforce, Gmail) handles the information ChatGPT Work reads from it.
The New Security Surface Area
This is the detail most coverage of the launch has glossed over. A chatbot that only responds to what you type has a fairly contained blast radius. An agent with standing access to your email, calendar, and internal chat tools is a fundamentally different risk profile — it can read things you didn’t explicitly hand it, and in the case of Scheduled Tasks, it can act without you actively supervising the session.
Before a broad rollout, IT and security teams should be asking:
- What exactly does each connected MCP plugin have permission to read versus write?
- Can actions be scoped per user or per team, or is access all-or-nothing?
- What audit trail exists for actions the agent takes autonomously?
- Does Scheduled Tasks require human approval for anything sensitive, or does it act independently by default?
- How does data flow between MCP-connected apps and OpenAI’s infrastructure, and where does it live at rest?
Compliance
Coverage of the launch references standard enterprise compliance credentials — SOC 2 Type II and ISO 27001-family certifications are typically cited by OpenAI for its enterprise products, along with GDPR controls and HIPAA business associate agreements where applicable. Data residency options across multiple global regions are also part of the enterprise offering. Organizations with strict regulatory requirements should confirm current certification status directly with OpenAI rather than relying on general product marketing, since compliance scope can vary by feature and by region.
Real-World Results: Does It Actually Work?
Customer Success Stories
OpenAI has publicized a handful of early results. Zapier reported recovering a seven-figure sales pipeline using the tool. Internally, OpenAI’s own sales team says it can now go from initial discovery to a proof-of-concept in about 24 hours, down from what used to take weeks, and its finance team says month-end close has moved from a multi-day process to a matter of hours.
OpenAI’s Internal Adoption
OpenAI product manager Ty Geri, in an interview about the launch, said the tool has become central to how the internal product team operates — describing a case where ten separate bug-bash sessions were scheduled across the team in minutes rather than the back-and-forth that normally takes. Geri put the internal impact bluntly:
“My job is not to schedule bug bashes… my job is to make an amazing product.”
That’s the pitch in miniature: offload the coordination overhead, spend the freed-up time on the work that actually requires judgment.
