ChatGPT computer use agents now drive a real desktop — clicking, typing, and moving files between apps while you keep working — and the part most people misunderstand isn't that the AI can do this, it's that you don't pick a "computer use" tool at all. As of July 9, 2026, OpenAI merged three modes (Chat, Work, and Codex) into one ChatGPT desktop app on Mac and Windows, retired the standalone Codex app, and folded its Atlas browser into a built-in alternative powered by GPT-5.6 Sol, Terra, and Luna. The short version: Chat is a conversation, Work is a long-running knowledge-work agent that hands you finished files, and Codex is the same agent pointed at code, PRs, and terminals (OpenAI Help Center). Treat anything that runs on a website as untrusted, confirm the active account before a click, and start with low-stakes tasks — these are the guardrails that turn a flashy demo into dependable output.
Last verified: 2026-07-26 — The desktop app is free to try on Mac and Windows; paid plans unlock longer runs, web/mobile use, and the full GPT-5.6 family — last checked July 2026. Pricing and limits change often.
TL;DR
- The new ChatGPT desktop app unifies Chat, Work, and Codex; the old desktop app is now "ChatGPT Classic" and still works.
- Work is the headline feature: a knowledge-work agent that produces finished spreadsheets, decks, documents, dashboards, and even shareable web apps via the Sites beta (OpenAI).
- Computer use runs in the background — one-time or scheduled — with a permission prompt before every new website or app.
- GPT-5.6 ships as Sol ($5/$30 per 1M tokens), Terra ($2.50/$15), and Luna ($1/$6); the same family powers all three modes (9to5Mac).
- Codex is now a developer view inside the same app, with inline diff editing, sidebar PR review, and faster computer use.
- The standalone Atlas browser is being retired on August 9, 2026; its agentic browsing is folded into the new built-in browser (The Verge).
What is a ChatGPT computer use agent?
A ChatGPT computer use agent is an AI that controls a browser or a desktop app on your behalf — visiting sites, clicking buttons, filling forms, downloading and creating files — instead of only answering a prompt in a chat window. OpenAI first previewed this as the Computer-Using Agent (CUA) behind Operator in January 2025; Operator was retired in August 2025 and its capabilities merged into ChatGPT Agent, which is now folded into the Work and Codex modes of the unified desktop app (OpenAI, OpenAI Help Center). The agent runs in its own virtual computer or on your local desktop with your permission; you don't watch every click — you approve the plan, then keep working.
What changed on July 9, 2026?
OpenAI shipped three things on the same day:
- The new ChatGPT desktop app for Mac and Windows — which now contains Chat, Work, and Codex as separate modes under one menu. Existing Codex app users can update in place and keep their projects, settings, and workflows. The previous ChatGPT desktop app was renamed ChatGPT Classic and is still available (OpenAI Help Center, NodeMini).
- ChatGPT Work — a new agent mode for long-running knowledge work. The agent can plan, connect to Gmail, Slack, Google Drive, SharePoint, Salesforce, and other tools via
@-mentions, and run for hours to deliver finished spreadsheets, decks, and documents (PromptsRush). OpenAI also quietly released a Sites (beta) feature so Work can publish a result as a live, shareable web app — dashboards, project trackers, launch calendars — that stays current as the underlying data changes (OpenAI). - GPT-5.6 Sol, Terra, and Luna — released for general availability after a two-week limited preview. Sol is the flagship for the hardest work; Terra is balanced on price and capability; Luna is the fastest and cheapest. All three power Work and Codex depending on your plan (OpenAI).
On the same day, OpenAI confirmed the standalone ChatGPT Atlas browser will stop working on August 9, 2026, with its agentic browsing folded into the new built-in in-app browser and a separate cloud browser for remote agent work (The Verge, Piunika).
How do Chat, Work, and Codex differ?
The three modes share the same AI backbone and computer use engine, but they are aimed at different audiences and deliverables. Picking the right mode is the most important early decision; using the wrong one wastes token quota and produces the wrong shape of output.
| Mode | Best for | What it produces | Who gets it |
|---|---|---|---|
| Chat | Everyday questions, drafting, quick lookups | Conversational answers, short artifacts | All plans, free included |
| Work | Long-running knowledge-work projects that need files | Finished spreadsheets, decks, docs, dashboards, shareable Sites | All plans on desktop (limited on Free) |
| Codex | Developers — code, repositories, PRs, terminals | Code edits, PR reviews, multi-repo changes | All plans on desktop |
Source: OpenAI Help Center and NodeMini — confirmed July 9–10, 2026.
A practical way to think about it: Chat is what to use when you want an answer; Work is what to use when you want a finished thing; Codex is what to use when the finished thing is code. The integration layer reaches the same app catalog in Work and Codex (Slack, Gmail, Drive, etc.), so the difference is who the deliverable is for — a colleague, a customer, or another developer's repo.
GPT-5.6 Sol, Terra, and Luna: which model do you use?
The model family split is the second decision. Each tier has its own price and stretch of capabilities, and max or ultra reasoning intensities stack on top of the tier choice. OpenAI's stated recommendation is choose a tier first, then increase reasoning only if the task needs it (knightli.com).
| Model | API price (per 1M tokens) | Best use case | Reported benchmark notes |
|---|---|---|---|
| GPT-5.6 Sol | $5 input / $30 output | Hard coding, long-running agents, research, cybersecurity | Artificial Analysis Intelligence Index v4.1: 59 at max — one point below Claude Fable 5 (with fallback) at ~⅓ the cost (AIToolsReview) |
| GPT-5.6 Terra | $2.50 input / $15 output | Everyday chat, office documents, routine programming | Balanced; widely available in Work and Codex for paid plans |
| GPT-5.6 Luna | $1 input / $6 output | Classification, summarization, extraction, batch processing | Fastest and cheapest; ideal for preprocessing before escalating to Sol |
Source pricing: OpenAI and 9to5Mac. Independent benchmark numbers from Artificial Analysis, published July 9, 2026. Independent evaluator METR reportedly flagged unusually unreliable capability numbers for Sol at release — treat any vendor-published benchmark with care and run a small task yourself before committing production work (AIToolsReview).
A sensible routing pattern, borrowed from how teams route traffic across model tiers: send classification and extraction to Luna; default everyday knowledge work and coding to Terra; escalate low-confidence, repeatedly-retried, or genuinely hard tasks to Sol. This keeps the expensive tier focused on cases where its judgment is actually needed — see our single-agent vs multi-agent decision framework for how this tiered pattern fits inside a larger automation stack.
How do you turn on computer use in the ChatGPT desktop app?
You don't toggle a "computer use" feature directly. You pick a mode, describe a task, and let the agent reach for a browser or local app only when it needs to. Here is what actually works in July 2026.
- Download the new desktop app from chatgpt.com/download for Mac or Windows. If you already had the Codex app, accept the in-app "Update" — your projects and settings are preserved and the app becomes the unified ChatGPT client (ScapesNews).
- Sign in with the same ChatGPT account you use on web. Free-tier accounts can try Work and Codex on desktop with usage limits; paid plans get bigger quotas and access on web/mobile.
- Switch to Work or Codex from the top-left menu. Chat remains the default for conversations.
- Connect tools (Work) by typing
@plus the tool name — for example@Gmailor@Google Drive— to pull an integration into the task. On Codex, open a repository or a project folder. - Describe the task in plain language with a clear finish line. Something like "Pull last month's invoice PDFs from @Google Drive, build a comparison spreadsheet with amount and due date, and flag anything overdue" reads like one sentence and runs like a workflow.
- Open the built-in browser when the task needs the web. The keyboard shortcut is Cmd+Shift+B on Mac or Ctrl+Shift+B on Windows. The agent can see the page you're on, move across multiple tabs, sign into sites, and download files — all without leaving the app (OpenAI Help Center).
- Review the plan, approve, and let it run. Watch the permission prompts the first time the agent touches a new site or a local app — every new domain asks before it runs, and you should treat anything on a website as untrusted information.
The Chrome extension is the alternative if you already live in Chrome all day. It lets ChatGPT work inside your real sidebar using your existing signed-in sessions, open tabs, and installed extensions — so you don't have to log into every site twice inside the built-in browser (OpenAI Help Center).
A note for Windows users: the in-place Codex-to-ChatGPT migration appears reliable on macOS, but at least one Windows reviewer hit a "Windows setup didn't finish" loop that survived reinstalls, administrator runs, and a fresh download. If you hit this, it is being reported; for now, the macOS path is smoother (ScapesNews).
What can ChatGPT Work actually do right now?
The headline capability is finished artifacts, not drafts. Work produces a deliverable you can hand to a colleague — formatted spreadsheets with working formulas, presentation decks, polished documents, live dashboards, and via the Sites beta, a shareable web app for the whole team (PromptsRush). What makes that useful in practice is that the output is the artifact itself, not a chat answer you have to copy out and reformat.
Concretely, Work today can:
- Read and assemble research across multiple browser tabs and connected tools (Gmail, Drive, Slack, Teams) into one document.
- Build spreadsheets and dashboards from connected data sources — and keep them updated when the underlying data changes.
- Produce internal Sites: dashboards, project trackers, launch calendars, and lightweight internal portals — hosted by OpenAI and shared by link (Business and Enterprise plans; not in Plus — see the OpenAI Sites landing page).
- Use Annotations to point at a specific spot on a page, document, or app and ask for a change there, instead of re-prompting from scratch.
- Run as a one-time action (do this once) or a scheduled task that repeats on a timer or when something changes.
The ceiling matters as much as the headlines. Sites is in preview and OpenAI-hosted — your tool lives on OpenAI's infrastructure, on a feature that is days old. Match the stakes to the maturity: internal dashboards yes, customer-facing checkout flows no. Always confirm sharing settings before sending the URL, especially for anything with customer or financial data.
What can Codex do as a developer mode?
Codex is now a view inside the ChatGPT desktop app — not a separate product — and it kept the developer-facing upgrades the standalone Codex app gained through 2026. As of July 2026 it can:
- Run computer use across local desktop apps, with inline diff editing inside code differences.
- Review GitHub pull requests directly in the sidebar, with reviewer feedback alongside the diff, without leaving the app (9to5Mac).
- Operate across multiple repositories in a single Codex project.
- SSH to remote devboxes (labeled alpha) — useful when the long agent runs benefit from stable compute and runtime away from a laptop (SmartScope).
OpenAI reports more than 5 million weekly Codex users, with the fastest-growing slice made up of non-developer knowledge workers — a sign Work and Codex are converging rather than staying cleanly separated (Sayfe.ai).
How does the permission system work, and what should you watch for?
OpenAI built a permission system on purpose, and it is the second thing — after the model tier choice — that decides whether computer use is dependable. Every time the agent wants to use a site or app it has not touched before, it asks first, unless you have pre-authorized that source (OpenAI Help Center). Treat anything on a website as untrusted information, and check which account is active before letting it continue — if it somehow opens the wrong account or the wrong page, stop it there.
Five working rules, distilled from primary-source docs and our own notes on perception agents and computer use reliability:
- Confirm the active account before each consequential click. A wrong login session is the most common silent failure.
- Use take-over mode for logins instead of sharing credentials directly with the agent.
- Avoid payment workflows in the first month — they're the highest-stakes test and the easiest to misjudge.
- End every task with "list anything you were unsure about." That one line surfaces the rows a human needs to spot-check, which is what turns polished output into trustworthy output.
- Match stakes to maturity. Internal tools first, customer-facing flows later, and never put your only copy of critical data inside a preview feature like Sites.
What does this mean for you?
For most people — including the small business and side-project audience we write for — the right move in July 2026 is to download the new desktop app, switch to Work mode on a low-stakes task, and pick the model tier that matches the budget. The honest framing is this: ChatGPT computer use agents are past the demo phase for scoped, well-framed tasks — research roundups, price checks across a dozen sites, first-draft spreadsheets — but they are still the wrong tool for judgment calls, sensitive accounts, and final drafts. The path from a flashy demo to a dependable workflow is the framing of the task itself: explicit deliverable, named data sources, named guardrails, and a small first run you can verify.
If you want to set up an actual agent pipeline instead of one-off tasks, our small-business AI automation stack guide walks through the delegation pattern and the stack that replaced a $20K/month support team. If you want the bigger picture on how multi-repo, multi-model orchestration fits around this, read our AI agent OS builder's guide. And if you want to turn a transcript or research doc into something you publish — the same workflow that produced this piece — our NotebookLM-as-research-engine teardown lays out the four-step pipeline.
FAQ
Q: Is ChatGPT computer use free to try? A: Yes. The new ChatGPT desktop app is available to download for free on Mac and Windows, and all three modes — Chat, Work, and Codex — can be used on the free tier with usage limits. The full GPT-5.6 family and longer runs are gated to paid plans (Plus, Pro, Business, Enterprise) (OpenAI Help Center).
Q: What is the difference between ChatGPT Work and ChatGPT Agent? A: Work is the new project-scale agent launched July 9, 2026 — multi-hour, multi-app, multi-artifact. The earlier "ChatGPT agent" mode launched in 2025 as a single browsing-and-forms session that merged Operator's CUA capability into a chat. Work is the broader productization of that capability across connected tools and finished deliverables, not just single browsing sessions (PromptsRush).
Q: What happens to users of the old Codex desktop app? A: Nothing breaks. When you update the standalone Codex app after July 9, 2026, it upgrades in place to the new ChatGPT desktop client and your projects, settings, and workflows are preserved. You can keep Codex as your default view and even keep the Codex app icon on macOS (ScapesNews).
Q: Do I need a separate subscription for ChatGPT Work? A: No. Work is included in existing ChatGPT plans on desktop. It uses the same usage-based billing as Codex, so longer, more complex tasks consume more quota. OpenAI has not published per-unit pricing; run a small test task first to gauge real consumption (NodeMini).
Q: When will the ChatGPT Atlas browser stop working? A: August 9, 2026. Atlas launched in October 2025 and is being retired about nine months later; its agentic browsing capabilities are folded into the new built-in browser in the ChatGPT desktop app and a separate cloud browser for remote agent tasks. Export any browser data you want to keep before then (The Verge, OpenAI Help Center).
Q: Which GPT-5.6 model should I pick for a small-business task? A: Default to Terra for everyday work — it's balanced on cost and capability at $2.50/$15 per 1M tokens. Use Luna ($1/$6) for classification, extraction, or batch work where you can verify the output mechanically, and escalate only the genuinely hard, repeatedly-failed tasks to Sol ($5/$30). The tier-first routing pattern keeps the bill under control without losing quality on the cases that matter (OpenAI, knightli.com).
Discussion
0 comments