Manus

There are two things to consider: which tasks you will hand it and whether you are ready to pay credits for each one.

Who it is for. For those handing off long tasks end-to-end: sort through a stack of resumes, turn data into analytics, build a website or slides by morning. Manus takes these tasks to a finished file itself, without your guidance along the way.

Who it is not for. Anyone expecting quick answers to questions: regular chatbots are for that, while Manus works in minutes and charges hundreds of credits per run.

Paused until August 25. The service is unavailable on 23–24 August: after the split from Meta, the vendor is deleting data created since Dec 29, 2025 and promises to restore access on Aug 25, 2026 at 8:00 Singapore time. New users can start on the 25th.

It's a fit. Section 3 will take you from registration to your first task. Choosing between it and ChatGPT Agent: the breakdown is in section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

One cycle for any task: a plan, work in its own virtual machine, and a finished artifact at the end.

Task in words. “Build a quadrant chart of NBA player performance” from the official demo is the entire input. The agent then creates a plan, finds data, writes and runs code in an isolated sandbox, and checks the result itself.

Work is visible. Next to the chat is the “Manus’s Computer” panel: it shows which command the agent is running, which website it opened, and which file it is editing. If it goes off track, correct it in words as it works.

Wide Research. For large tasks, Manus deploys up to a hundred copies of itself in parallel. Vendor examples: rank 100 MBA programs or analyze 1,000 stocks. Each copy works as a full agent, not a narrow assistant.

Browser Operator. The agent operates a browser like a person: logs in, clicks, fills out forms, and pulls data from live websites. The feature has been available to all users since November 22, 2025.

What was actually done. In the official demos: screening a pile of RL engineer resumes with a final ranking (launch Mar 6, 2025) and an NBA quadrant chart with player avatars. In third-party tests: a wedding invitation site in 25 minutes and 360 credits, a weather forecast web app for roughly 900 credits (cyber.sports.ru, June 2025). We have not run Manus ourselves: during our check window the service was paused, so the numbers come from the official demos and third-party tests.

There are three doors: browser, phone apps, and desktop. First, two preparation steps.

Step 1. Register at manus.im: email, Google, Apple, or Microsoft, with no phone number required. An invite code is not needed either — since 2026, it has been a 500-credit referral bonus rather than a pass.

Step 2. Give it your first task and name the outcome: “build a one-page invitation website,” “turn these tables into a report with charts.” The free 300 credits per day are enough for one or two small tasks — you can get a feel for the agent without paying.

Apps. The web app works in any browser. The iOS app is in the App Store, Android is on Google Play, and the site offers desktop apps for macOS and Windows. Don't install APKs from third-party sites: the official build is on Google Play.

Steps checked August 23, 2026 using vc.ru and Bitrix24 guides · service paused until August 25, 2026 08:00 SGT

Benchmark pairs of “agent + model” are not yet measured for general-purpose agents — look at three other signals.

Vendor claim. At launch, Manus claimed a GAIA benchmark record: 86.5% at level one versus 74.3% for OpenAI Deep Research (claim dated March 6, 2025). There is no independent reproduction of this figure, and GAIA itself is criticized as easy to optimize for — treat it as advertising, not a measurement.

Money as a measurement. Annualized revenue grew from $125 million in December 2025 to $450 million by June 2026 (Sacra estimates). Paid demand tripled in six months—even as Meta and regulators spent that entire time dividing the company.

You do not pay for access, but for work: every task consumes credits, and that is the main thing to understand before subscribing.

Free entry. 300 credits per day, expiring after 24 hours and not accumulating. Users spent 360 credits on an invitation website and about 900 on a web app with live data, so a free day is enough for one small task or a couple of very small ones.

Paid tiers. From $20 per month for 4,000 credits to $200 for 40,000; all paid plans retain the daily 300. Official pricing was unavailable during the review window along with the website, and the figures are from June 2026 review summaries, so they may change after the August 25 relaunch.

Three questions: where your data lives, what can happen to it, and how much to trust the results.

Data lives with the vendor. All tasks run in the Manus cloud, and August 2026 showed what that means in practice: at regulators' request, the vendor deleted user data created after 12/29/2025 in several jurisdictions. The lesson is simple: download finished artifacts immediately; Manus conveniently provides them as files.

Which jurisdiction. The company was founded in China, while its legal entity and headquarters are in Singapore; the vendor says its data storage is in the US and Singapore. The Chinese regulator blocked the Meta deal. A Tencent-led consortium is now negotiating to buy Meta's stake—the deal has not closed, and the ownership structure is still in flux.

Results need fact-checking. Reviewers agree: the agent "hides hallucinations well," and its report looks polished even when the numbers are wrong. Long tasks can loop; a known failure was booking a trip to Japan. The agent runs code in an isolated sandbox and does not touch your machine.

The choice is not "who is smarter," but "what work and what budget."

If you need to compare text and image variants side by side, see the Flowith workspace canvas: prompts and results are linked in branches.

If you need a web task runner inside your own service, compare the current ways into H via the Runner H card: it separates the archived Studio from the managed Agents API.

Different economics. ChatGPT Agent is included in the $20 Plus subscription with no credit counter: practitioners call it the safest default choice for that reason. Manus charges credits for every run, and a complex task costs 300–900 - for regular work, it is more expensive.

Different depth. Manus handles long autonomous tasks and delivers finished artifacts: a website, slides, a report with charts; Wide Research can spin up to a hundred agent copies for one task. ChatGPT Agent is faster and more convenient for small tasks within the familiar chat.

Who else is nearby. Genspark is faster on presentations and multimedia, the Comet browser builds the agent right into the browser.

For summaries of your work tabs, email and documents, consider the Dia browser.

If you want to connect an agent of your choice to your tabs via MCP or CLI, consider the Opera Neon browser.

The product is a year and a half old, and almost all of its second year was spent deciding who it would belong to: the Meta deal, the regulator's ban, the relaunch.