Osum · Updated
ChatGPT Work vs Manus vs Comet (2026)
The demo is easy. The second week is where agents win or lose your month. That changes how you should compare ChatGPT Work, Manus and Comet: all three can act for you, but they differ on who plans the run, where the work lives and who checks it before it leaves. Here is how the three compare on setup, supervision and the jobs each one finishes in week two.
The short answer
Pick ChatGPT Work if you want an agent inside the ChatGPT you already pay for, working from your files and tools, with a plan you approve before anything runs. Pick Manus if you want a general agent that researches at scale and builds and deploys whole apps, and you are happy to learn its credit system. Pick Comet if you want the agent living in your browser, acting on the pages in front of you, rather than running tasks from a chat.
Users who stick with agent work describe the same pattern: watched and scheduled jobs succeed, and open-ended “handle everything” requests stall and burn usage. Start with a job you can check against a source, then hand over more.
| Task | What users report | Your decision |
|---|---|---|
| Price and deal watching | Alerts and monitored pages work well | Start with one page and an exact condition. |
| “Handle everything” requests | Messy behaviour and burned usage | Write the four-part brief first, then approve the plan. |
| Half-finished runs | Thirty minutes for two entries, then a suggestion to start over | Keep a recovery record and resume the unfinished step. |
| Client-facing deliverables | Polished pages with unsupported claims | Check every claim against its source before it leaves. |
What's new in ChatGPT Work for 2026
OpenAI retired the agent name. Its help center now says ChatGPT agent “is no longer available” and points longer, multi-step work to ChatGPT Work; Operator was folded in earlier and its site is no longer accessible (both checked 28 September 2026). If you learned the old mode, the skills carry over; the surface is now Work, chosen from the toggle in ChatGPT.
Work and Codex now share one usage pool: Work “follows the same usage structure as Codex”, with usage that varies by task and credits to extend it. Scheduled tasks can run once, repeat on a schedule or trigger, or monitor for changes, and eligible paid plans can run event-triggered tasks on new Gmail messages, new Slack messages and GitHub pull-request activity.
One dated change to plan around: GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on all plans on October 14, 2026. If you saved a method tuned to it, re-run your checks after the switch.
What the rename changes for the pick: any comparison written when “agent mode” was the product is now answering a page that no longer exists. Judge Work on a planned, approved run and on the second week of a schedule, not on the first demo.
What ChatGPT Work is and how it differs from ChatGPT
Chat answers questions and drafts text. Work “is an agent designed for longer, multi-step work and finished deliverables”: it gathers context, plans the approach, and acts across your tools, files and desktop apps to produce spreadsheets, docs, slides and Sites. Codex stays dedicated to software development.
The plan comes before the action. In Plan mode, Work gathers context, asks questions and creates a step-by-step plan; you suggest changes or approve the plan to start the work. Work runs on web and mobile on paid plans and in the desktop app, where it can also use local files with your permission.
ChatGPT Work vs Manus, Comet and Claude Cowork
Here is the practical difference. ChatGPT Work is a feature of the ChatGPT plan you may already pay for, running tasks with the context and apps you connect to it. Manus is a separate general agent with its own credit system, built around wide research and building complete apps. Comet is Perplexity’s browser, where the agent acts on the pages you browse. Claude Cowork covers the desk-work slice: finished documents from your own files.
The plan-first problem follows you to every one of them. Whichever agent you pick, someone has to say what done looks like before the run starts, and someone has to check the result before it leaves. That job is yours on all four.
| Option | Where it runs | Paid from | Best at | Watch out for |
|---|---|---|---|---|
| ChatGPT Work | ChatGPT web, mobile and desktop app | Free tier on desktop; Plus was $20 a month when checked on 28 September 2026 | Multi-step work and finished deliverables from your connected tools and files | Usage is shared with Codex and varies by task, so open-ended requests burn it |
| Manus | Manus web app, browser extension and mobile apps | Free to start (checked 28 September 2026) | Wide research across many sources, and building and deploying complete apps | Its own credit system and sandbox to learn; another plan to manage |
| Comet (Perplexity) | A browser for Mac, Windows, iOS and Android | Free to download (checked 28 September 2026) | Acting on the pages you browse, with research answers beside them | You browse inside Perplexity’s browser; the agent confirms actions before it proceeds |
| Claude Cowork (Osum course) | The Claude desktop app | Needs a Claude plan, billed separately | Finished reports and documents from your own files | Desk documents, not web runs |
ChatGPT Work vs Manus: who plans the run
Manus is built to do more on its own. Its own comparison page leads with wide research across a hundred or more sources in parallel subagents, apps built and deployed end to end, and recurring tasks that run while you are away, all in its own cloud sandbox with a file system and terminal. If your job is “research everything and build the tool”, Manus starts further along.
ChatGPT Work starts from your context instead: the apps, files and approvals already attached to your ChatGPT plan, with a plan you approve before the run starts. That approval step is the difference that decides week two. On either agent, an unstated step becomes a guess; on Work, the guess is caught on the plan, before it costs usage.
Choose Manus if you want maximum autonomy and are willing to manage its credits and sandbox as a second tool. Choose ChatGPT Work if you want agent work inside the ChatGPT plan you already run, and a plan on your desk before anything acts.
ChatGPT Work vs Comet: the chat or the browser
Comet puts the agent in your browser. Perplexity’s page describes a personal assistant that researches, drafts email replies, shops and builds sites while you browse, with a clear control model: Comet asks you to confirm actions before it proceeds, you can stop it at any time, and Assistant and Agent are separate modes.
ChatGPT Work runs the job from a chat instead, on a plan you approve, drawing on your connected apps and files and finishing as a deliverable: a spreadsheet, a doc, a deck or a Site. The trade is context: Comet sees the browsing session you are in; Work sees the workspace you connected.
Choose Comet if your work lives in pages you are already looking at, and you want the agent one keystroke away while you browse. Choose ChatGPT Work if the job starts from your files and tools and ends in a finished deliverable, with the plan approved before anything runs.
What ChatGPT Work costs right now
There is no separate ChatGPT Work subscription. Work is part of ChatGPT plans: the desktop app includes it on every plan, including Free, and web and mobile need Plus, Pro, Business, Enterprise or Edu.
Work’s usage comes from the same pool as Codex. Usage varies with the task, so two jobs that look similar can cost different amounts of your allowance, and credits can extend it. Check the usage dashboard in your plan before you hand over a large job, and remember that a stalled, retyped run costs the same as a finished one.
One date to note: GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on all plans on October 14, 2026.
These plan prices were checked on September 28, 2026, on OpenAI’s pricing page.
| ChatGPT plan | Monthly price | ChatGPT Work access |
|---|---|---|
| Free | $0 | Desktop app |
| Go | $8 | Desktop app |
| Plus | $20 | Web, mobile and desktop; shares usage with Codex |
| Pro | From $100 | Web, mobile and desktop, with 5x or 20x the Plus usage |
| Business | $20 per user, billed annually ($25 monthly) | Workspace controls for apps and features |
Where agent runs break after the demo
A seller gave agent mode a straightforward job: fill twenty-five fields in a WordPress form, correctly. It ran for a month’s worth of usage on that one task. The failure was not the typing; it was a job shaped like one prompt instead of a series of checks nobody wrote down.
The stall is the second pattern. One user watched a run put in two entries in thirty minutes, announce it had run out of context, and suggest starting again in a fresh session. Retyping the whole job would have paid for the same work twice. What survives the stall is a record of what finished, kept outside the chat.
The takeover is the third. A user mid-login watched the takeover window vanish on eBay and other platforms, which matches OpenAI’s own note that a run sometimes needs to re-establish itself after you hand the browser back. Keep the sensitive step short, and never type a password into chat.
Each pattern has the same fix in one sentence: decide what done looks like before the run starts, keep the record in a project rather than a chat, and check the deliverable against its source before it leaves.
ChatGPT Work FAQ
Is ChatGPT agent mode gone? Yes. OpenAI’s help center says ChatGPT agent “is no longer available” and points longer, multi-step work to ChatGPT Work (checked 28 September 2026). Old how-to pages still describe the retired mode, so check the date on anything you read.
Does ChatGPT Work cost extra? No separate subscription: Work is included in ChatGPT plans. The desktop app includes it on every plan, and web and mobile need Plus or above. Plus was $20 a month when we checked on 28 September 2026.
How does Work usage work? Work and Codex share one usage pool. Usage varies by task, so a stalled, retyped run costs the same as a finished one, and credits can extend the allowance.
Can Work use my files? Yes, in the desktop app, with your permission, one folder at a time. Work on web and mobile cannot read your computer directly, and cloud chats sync across devices.
Can Work log in to sites for me? It pauses and hands you the browser with takeover. Screenshots pause while you hold the controls, so what you type stays out of the transcript, and some sites still refuse agents entirely.
Can it run every week without me? Yes: tasks run once, repeat on a schedule or trigger, or monitor for changes, managed at chatgpt.com/schedules. Each run uses your plan’s usage, so a scheduled job is worth briefing as carefully as a manual one.
What will it refuse? High-risk tasks such as bank transfers are refused by design. Money and identity steps stay with you.
Verdict: which agent should you use?
ChatGPT Work is the pick if you want agent work inside the ChatGPT plan you already have, on your own files and tools, with the plan approved before anything runs. Choose Manus if you want a standalone general agent that researches at scale and ships whole apps, and choose Comet if the agent should live in your browser. If the job is documents from your own files, the Claude Cowork course covers that surface.
Whichever you pick, the skill that decides whether it saves you time is the same: scoping the job, reading the plan, approving it, and checking the deliverable against its source. We call it the Approved Run. If you’re going with ChatGPT Work, our course teaches it in 12 hands-on lessons, with a certificate and every other Osum course included.