Codex merged into desktop · Chat/Work/Codex modes · 1,400+ plugins · Plan Mode · vs Claude Cowork
If you are a knowledge worker, developer, or team lead evaluating AI work agents, OpenAI's July 9, 2026 launch of GPT-5.6, ChatGPT Work, and the merger of the standalone Codex app into a new ChatGPT desktop app is a product-level shift—not a routine feature drop. This guide delivers the three-mode desktop overview, five core Work capabilities, four Codex upgrades, a full comparison table vs Claude Cowork, usage-metered pricing, a six-step getting-started runbook, and seven FAQ answers—so you can pick the right agent route for browser SaaS workflows. Last updated: 2026-07-10
On July 9, 2026, OpenAI rolled out GPT-5.6 and quietly executed a strategic pivot: the standalone Codex desktop app was officially discontinued, with all capabilities moved into a new ChatGPT desktop app, alongside the launch of ChatGPT Work—an AI agent for knowledge workers. This is not "Codex being killed." It is OpenAI's push toward a super app.
Starting July 9, 2026, the standalone Codex desktop app is no longer distributed separately. Existing users do not need a fresh install—just update the Codex app and it automatically becomes the new ChatGPT desktop app. Projects, settings, and workflows are preserved. Developers can set Codex as the default launch mode and even keep the Codex app icon on macOS, so the transition feels seamless. Desktop Codex projects remain accessible from the ChatGPT mobile app.
The previous ChatGPT desktop app was renamed ChatGPT Classic and remains available for users who want a pure chat experience.
| Mode | Purpose | Best for |
|---|---|---|
| Chat | Everyday Q&A and conversation | All users |
| Work | Cross-app autonomous tasks with finished deliverables | Professionals and knowledge workers |
| Codex | Professional coding agent: reviews, PR management | Developers and engineering teams |
Free tier access: All users, including the free plan, can access all three modes on desktop. Web Work access is limited for free users—the desktop app is the primary entry point for free-tier Work capabilities.
For teams choosing an AI agent, these five pain points are worth unpacking early:
"Codex is gone" misread: Codex as a standalone app is discontinued, but as a dedicated desktop mode it remains—now with inline diff editing, PR side panels, and more. Do not confuse "app merger" with "feature removal."
Chat / Work / Codex boundary blur: All three share the plugins directory, but Work hides implementation details for non-technical users while Codex exposes more technical context—picking the wrong mode creates a real experience gap.
Hidden usage-metered costs: Work and Codex share the same usage-metered structure. Multi-hour cross-app tasks may burn more quota than a single chat session. OpenAI has not published per-task rates.
Web vs desktop capability gap: Computer Use and local file control are desktop-only. Free users face Work limits on web and mobile—teams must decide whether their main battlefield is browser SaaS or local folders.
Philosophy clash with Claude Cowork: OpenAI bets on cloud + desktop hybrid with 1,400+ plugin breadth; Anthropic bets on local folder sandboxes + native M365. Running both agents in one team can mean duplicate spend and split workflows.
ChatGPT Work is the headline product of this release—OpenAI's "AI coworker." You state a goal; it works autonomously across apps for hours and delivers finished outputs, not just suggestions or outlines. A typical flow: ① draft an execution plan and wait for your approval → ② connect Slack, Gmail, Google Drive, and other tools to gather context → ③ complete multi-step work independently → ④ deliver a document, spreadsheet, presentation, or web app.
ChatGPT Work connects to mainstream work tools through a unified plugins directory. Launch coverage includes:
Usage is simple: type @AppName in your prompt to pull from a specific source, or describe the task and let the AI choose the right integration.
For complex tasks, Work introduces Plan Mode: the AI lists execution steps first; you review and approve before work begins. This prevents drift and keeps you in control of high-stakes workflows.
On desktop, ChatGPT Work has Computer Use capabilities: read and edit local files, browse the web in a built-in multi-tab browser, click, type, and move files on your behalf, and run one-off tasks or Scheduled Tasks.
Set tasks to run on a schedule, when a trigger fires, or on a recurring basis—even when you are not at your desk, the agent keeps advancing the work. OpenAI's framing: Work can stay with a project for hours, turning goals into finished work—addressing the old Agent mode limitation of stopping after a few minutes.
The competitive focus is shifting from "whose model scores highest on benchmarks" to "whose agent is most deeply embedded in daily workflows"—ChatGPT Work's answer: on your desktop, in your Slack, in your Google Drive.
Inline diff editing: edit directly in the diff view without copy-pasting between panels.
PR side panel review: pull request review and reviewer feedback on one screen—no context switching.
Faster Computer Use: powered by GPT-5.6 with materially improved execution speed.
Multi-repo project support: one project can span multiple codebases; edit Markdown and code in-app with inline annotations.
Anthropic launched Claude Cowork in April 2026. Both target "AI work agents," but the design philosophies diverge clearly:
| Dimension | ChatGPT Work | Claude Cowork |
|---|---|---|
| Runtime environment | Cloud + desktop hybrid | Local desktop first |
| File access | Local files on desktop; upload mode on web | Direct control of a designated local folder (sandbox) |
| Integration ecosystem | 1,400+ plugins, broader coverage | 20+ official MCP connectors, native M365 |
| Best use cases | Cross-web-app and cloud-tool collaboration | File-heavy, repetitive document production |
| Non-technical UX | High (approachable interface) | Very high (hides all technical concepts) |
| Pricing model | Usage-metered (complexity drives consumption) | Seat-based (Pro from $20/month) |
| Free tier access | Yes — limited on desktop | No free tier |
| M365 native add-ins | Web-only (no native Word/Excel/PPT add-ins) | Yes — native Word/Excel/PPT |
| Scheduled tasks | Yes (via Tasks) | Yes (native UI) |
Work lives in browsers and SaaS tools → choose ChatGPT Work. Work lives in local folders with batch document production → choose Claude Cowork. Serious workflow teams in 2026 will likely run both.
These six steps cover desktop (recommended—all plans including Free) and web/mobile rollout timing. Follow them in order to complete your first Work task:
Download or update the desktop app: go to chatgpt.com/download for the new ChatGPT desktop app (Mac / Windows). If you already have the Codex app, updating it automatically becomes the new app.
Switch to Work mode: after launch, switch to Work in the top navigation. Developers can set Codex as the default view and keep the Codex icon.
Connect the plugins directory: in Settings, manage plugins and connect Gmail, Slack, Google Drive, Salesforce, and other tools you need.
Describe the task with Plan Mode: state your goal in natural language (optionally with @AppName), wait for the AI to list steps, review and edit, then confirm execution.
Watch usage and validate deliverables: run a small task with known complexity first, observe quota consumption, then decide whether to scale up or configure Scheduled Tasks.
Pick web/mobile by plan: Pro, Enterprise, and Edu users get web/mobile from July 9; Plus and Business roll out over the following days; Free users face web limits—use desktop as your primary surface, with mobile for monitoring and managing Work tasks.
vs Operator: Operator is a browser automation agent. ChatGPT Work is broader—plugin integrations with external apps, local file access on desktop, document-grade deliverables, and hours-long multi-step projects rather than single browser tasks. OpenAI is also gradually sunsetting the dedicated Atlas browser, shifting some web-native tasks to a Chrome extension.
ChatGPT Work is not sold as a standalone product. It is included in existing subscriptions but follows the same usage-metered structure as Codex: more complex, longer tasks consume more quota. OpenAI has not published per-task rates—run a small test task first.
| Plan | Monthly price (US) | ChatGPT Work access |
|---|---|---|
| Free | $0 | Limited on desktop |
| Go | $8 | Expanded on desktop |
| Plus | $20 | Desktop + web/mobile |
| Pro | $100–200 | Full access, highest limits |
| Business/Enterprise | Custom team pricing | Full access + Admin Console (usage controls, group limits, individual overrides, credits request flow) |
From "best model" to "deepest workflow embed": GPT-5.6 matters, but the strategic bet is who gets deepest into Slack, Drive, and desktop files—delivering finished work, not advice.
Codex audience expands beyond developers: roughly 5 million people use Codex weekly; over 1 million use it for non-coding work. After the merge, everyday professionals can reach similar power without terminal experience.
The AI super app thesis takes shape: one app holding Chat + Agent + Coding + file control + scheduled tasks + 1,400+ plugins—OpenAI building a moat through surface area, not model scores alone.
ChatGPT Work and Claude Cowork both excel at cloud and SaaS orchestration, but engineering teams needing Xcode signing chains, Metal local inference, and long-running iOS CI/CD cannot replace dedicated macOS build nodes with chat agents alone—virtualized macOS carries performance loss and EULA risk; laptops are unstable for 24/7 duty. For a more stable production environment suited to iOS CI/CD and AI agent automation, VpsMesh Mac Mini cloud rental is usually the better fit: physical Apple Silicon, root access, and predictable monthly cost—complementing ChatGPT Work's orchestration layer with a "cloud Mac runs builds + Work pulls Slack/Drive context" architecture. See Mac Mini M4 rental pricing and the help center for deployment steps.
The standalone Codex app is discontinued, but Codex lives on inside the new ChatGPT desktop app as a dedicated coding mode. Update your existing app—projects, settings, and workflows carry over. You can still access desktop Codex projects from the mobile app.
Yes, with limits. The desktop app gives free users limited access to ChatGPT Work. Full web and mobile access requires Plus or higher. For team build nodes alongside Work, see Mac Mini M4 rental pricing.
Regular ChatGPT responds in a single turn. ChatGPT Work is an autonomous agent built for long tasks, cross-app execution, and finished deliverables—supporting hours of autonomous runtime, 1,400+ plugin integrations, and Plan Mode. Standard Agent mode fits single-step, short tasks that often end within minutes.
OpenAI has not published per-task rates. Usage varies with complexity and follows the same metered structure as Codex. Run one test workflow with Plan Mode before scheduling recurring automation. Enterprise admins can set group and individual limits in the Admin Console.
Mobile rollout is phased: Pro, Enterprise, and Edu users first from July 9, then Plus and Business over the following days. You can monitor and manage Work tasks from the mobile app while the desktop handles heavy execution. Free users should rely on desktop for the full experience.
Operator is a browser-based automation agent. ChatGPT Work is broader: plugin integrations with external apps, local file access on desktop, document-grade deliverables, and hours-long multi-step projects—not just single browser tasks. OpenAI is also sunsetting the dedicated Atlas browser in favor of a Chrome extension for some web-native workflows.
Depends on your workflow. ChatGPT Work wins for cloud-first, multi-app breadth and free-tier desktop access. Claude Cowork wins for local folder workflows and native M365 add-ins. Many teams in 2026 will run both rather than replace one with the other. For 24/7 macOS CI nodes, see the help center.