ChatGPT Work supports Plus, Pro, Business, and Enterprise on web and mobile

• ChatGPT Work now spans paid web and mobile plans plus the desktop app: thsottiaux's availability post says it is in Plus, Pro, Business, and Enterprise, while OpenAI's Work page also lists Edu and all-plan desktop access.
• Sites is the breakout surface: ChatGPTapp's Sites post pitched games, small businesses, research, and team hubs, while omooretweets' game thread showed a hosted GeoGuessr-style app with ChatGPT auth and secure key storage.
• Work is built around browser and computer-use tasks, not only document generation: reach_vb's mobile example connected email and calendar context to restaurant research, and ChatGPTapp's desktop demo showed browser-driven work across apps.
• OpenAI patched the desktop shell after launch feedback: thsottiaux's desktop update added visible history, Projects, Chat and Work syncing, and clearer ChatGPT/Codex switching.
• Usage limits were loose during the rollout: thsottiaux's reset announcement reset paid-user limits for Codex and ChatGPT Work.

OpenAI's July 9 release notes bury the product map: Work creates finished docs, sheets, slides, reports, and Sites; Sites public publishing is in beta; Atlas is being retired as browser-based agentic work moves into ChatGPT. The Sites docs expose the fun implementation bits: D1/R2 storage, oai-authenticated-user-email headers, production deployment URLs, analytics without an SDK, and custom domains. The browser docs add the safety model: page content is untrusted, the built-in browser uses a separate profile, and sensitive actions require confirmation.

OpenAI's launch post defines Work as an agent that can operate across apps and files, stay with a project for hours, and turn a goal into finished work. The same post says Codex technology is built in, with more than 5 million weekly Codex users and more than 1 million using it outside software development.

The current surface area is concrete:

• Work: longer tasks across connected apps and files, with progress updates, user steering, and approvals, according to OpenAI's release notes.
• Outputs: documents, spreadsheets, presentations, reports, and Sites, per the same release notes.
• Plans: web and mobile rollout to paid plans except Free and Go in the release notes; thsottiaux's paid-subscriptions reply described Work as part of paid subscriptions.
• Desktop: Chat, Work, and Codex in one macOS and Windows app, with desktop Work able to use local files and apps with permission, according to OpenAI's launch post.
• Plugins: the App Directory became the Plugin Directory, and existing app connections were left unaffected, per OpenAI's release notes.

The consumer framing was also explicit: thsottiaux's weekend prompt asked what people were building, and the follow-up listed sites, email management, document summarization, and docs, sheets, and slides.

ChatGPTapp's launch examples made Sites feel less like a deployment feature and more like a new default output format:

• Build a game.
• Launch a small business.
• Make an OOO handoff.
• Show off photography.
• Publish research.
• Create a team hub.

omooretweets used Sites for a GeoGuessr-style game where a human competes against LLMs, then said Work walked through Google Maps and Street View API keys plus five OpenRouter model keys in the API-key follow-up. The public result was WorldModel, and omooretweets' budget note said GPT-5.6 helped set an API-usage budget.

Publishing was the missing piece for vibe coding. omooretweets' publishing note described one-click public sharing and custom-domain support after the app was built.

The Sites documentation reads like a small hosting platform manual:

• Publishing has two stages, save a version and deploy a version, and every deployment URL is a production deployment.
• Durable structured data maps to D1, file uploads map to R2, and .openai/hosting.json stores bindings such as d1 and r2.
• Public Sites can add Sign in with ChatGPT through /signin-with-chatgpt and /signout-with-chatgpt.
• Signed-in visitors arrive with oai-authenticated-user-email and optional oai-authenticated-user-full-name request headers.
• Hosted environment variables and secrets live in Site settings, not prompts, attached files, or .openai/hosting.json.
• Custom domains require user-managed DNS, and custom domains are not available in Enterprise workspaces at launch.

dkundel's identity-provider note showed the user-facing version of the same feature: ask Sites to add ChatGPT as the identity provider, and it wires up the sign-in flow.

The browser is the other half of Work's pitch. In reach_vb's plugin reply, email and calendar connections were described as universal plugins that show up across surfaces, while the browser handles browsing tasks.

OpenAI's browser docs add the operating model:

• Browser is available on the web and in the ChatGPT desktop app.
• The desktop browser uses a separate profile from the user's regular browser.
• @Browser invokes the built-in browser.
• Computer Use can open pages, click, type, inspect rendered state, take screenshots, and verify results.
• ChatGPT asks before using a website unless it was already allowed.
• Sensitive actions such as submitting information, purchases, permission changes, or deleting data require confirmation.
• File uploads cannot be automated in the built-in browser.

The docs also tell users to treat page content as untrusted context. That line matters because Work is explicitly crossing from answer generation into action execution.

OpenAI had to clean up the unified desktop app in public. thsottiaux said the first try was not quite right, then listed the fixes: conversation history and Projects in the sidebar, Chat and Work history synced across web, mobile, and desktop, an easier Chat/Work switcher, and no change to Codex mode.

OpenAI's July 16 desktop notes turned those bullets into a changelog: global ChatGPT/Codex switcher, unified Recents, Projects in the desktop app, cloud Work sync across devices, and local conversations staying on the computer. An OpenAI community thread later said the team had released an update and that cloud projects were syncing in the latest desktop app.

The naming still landed awkwardly. simonw's OpenAI Claw joke captured the category confusion, simonw's documentation complaint said Work was hard to understand because desktop and mobile presented it differently, and dkundel's Quick Chat reply separated lightweight ChatGPT access inside Codex from the deeper ChatGPT mode.

After the desktop update, thsottiaux's follow-up kept the product-feedback loop visible rather than hidden in a changelog.

The GPT-5.6 Sol rollout and Work rollout collided with quota pressure. The sequence started with thsottiaux's Sol post calling GPT-5.6 Sol an extremely good model, then moved into a paid-user reset for Codex and ChatGPT Work.

In thsottiaux's follow-up, the reset may also have touched other rate limits. OpenAI's launch post says Work usage follows the same structure as Codex, with more complex tasks using more of a plan's included usage.

The feedback program was aimed outside the developer bubble. thsottiaux asked for 50 to 100 ChatGPT Work users, explicitly said he was not looking for developers this time, and asked people to DM how they used Work and what OpenAI should fix.

The public asks clustered around continuity and control:

• Proactivity and email webhooks, from jxnlco.
• Continuous project-context enrichment and memory appends, from reach_vb.
• More resets, from ericmitchellai.
• More stable ChatGPT/Codex desktop connections, from MatthewBerman.
• New folders and projects from Codex Remote inside the ChatGPT app, from chrisbbh.
• Documentation for what Work is across desktop and mobile, from simonw.

The meta example was Work being asked to process its own feedback pipeline. In thsottiaux's dictated prompt, ChatGPT Work was told to scan thousands of Twitter DMs, keep only those mentioning ChatGPT Work, write rows into a spreadsheet, classify each user's work into 8 to 12 labels, and rate workflow sophistication for cohort selection.