Manus is an autonomous AI agent originally developed by Butterfly Effect, a startup founded in China and later headquartered in Singapore. It launched on March 6, 2025, and was acquired by Meta for approximately $2 billion in December 2025. Unlike chatbots that respond to individual prompts, Manus breaks complex goals into subtasks, delegates them to specialized sub-agents, and executes multi-step workflows autonomously in the background. It uses Claude and Qwen models under the hood, and has been integrated into Meta Ads Manager and Telegram.

Is Manus AI owned by Meta?

Yes. Meta acquired Manus in December 2025 for approximately $2 billion. Manus continues to operate as a standalone product at manus.im with its own subscription plans, while Meta integrates the technology into its platforms including Ads Manager, and plans integration with WhatsApp, Instagram, and Facebook. The acquisition is under regulatory review by Chinese authorities over potential technology export concerns.

How much does Manus AI cost?

Manus offers a free tier with 1,000 bonus credits at signup and over 300 free credits daily. Paid subscriptions provide higher usage limits. The standalone product at manus.im uses a credit-based system where tasks consume credits based on complexity. The Manus integration inside Meta Ads Manager is available at no additional cost to advertisers using the platform.

Is Manus AI reliable?

It depends on the use case. For research, data compilation, and report generation, early users report strong results. For the Meta Ads Manager integration specifically, media buyers have reported hallucination issues with outputs that aren't reliable enough to send to clients without manual review. The standalone Manus product has improved significantly since launch — average task completion time dropped from over 15 minutes to under 4 minutes — but autonomous agents still require human oversight for high-stakes work.

How is Manus AI different from OpenClaw and Claude Cowork?

Manus is a cloud-based autonomous agent that runs in its own virtual environment, focusing on background task execution and multi-step workflows. OpenClaw is a self-hosted, open-source agent that runs on your own hardware and connects to messaging apps. Claude Cowork is a desktop agent that works within folders on your computer with strong sandboxing. The key difference is that Manus is now backed by Meta and being embedded directly into platforms with billions of users, while OpenClaw and Cowork remain standalone tools you must set up yourself.

Manus AI: What Meta's $2B Agent Actually Does One Year Later

One year ago today -- March 6, 2025 -- a startup called Butterfly Effect launched Manus from Shenzhen, and the AI world temporarily lost its mind. The demo video hit a million views in under 20 hours. Invite codes sold for $7,000 on Chinese secondhand markets. The Discord hit 186,000 members in days. Jack Dorsey endorsed it publicly. The GAIA benchmark scores topped OpenAI's. Two million people joined the waitlist within a week.

Nine months later, Meta bought it for $2 billion.

And now, in March 2026, Manus is quietly showing up inside Meta Ads Manager, rolling out personal agents in Telegram, and positioning itself as the AI agent for the next billion users who will never set up OpenClaw or install Claude Cowork.

The question isn't whether Manus matters. The $2 billion check settled that. The question is whether it actually works -- and what it means that the most aggressively deployed AI agent in the world is now owned by the company that runs Facebook, Instagram, and WhatsApp.

What Manus actually is

Manus is an autonomous AI agent. You give it a goal, it breaks the goal into subtasks, delegates those subtasks to specialized sub-agents, executes them in a cloud-based virtual environment, and delivers finished output. No step-by-step prompting. No babysitting.

This is different from chatting with ChatGPT or Claude. Those are conversations -- you ask, they answer, you ask again. Manus is delegation. You describe an outcome, walk away, and come back to a completed deliverable: a research report, a spreadsheet, a website, a presentation, an itinerary with real prices and real availability.

Under the hood, Manus runs on a multi-agent architecture built on Claude and Alibaba's Qwen models. A planner agent decomposes your goal. Execution agents handle the actual work -- browsing the web, writing code, managing files, filling out forms. A verification layer checks the output. The whole thing runs in an isolated virtual computer in the cloud, and you can watch a real-time replay of exactly what it's doing -- which tabs it opens, what it searches, what it copies -- through a side panel that Manus calls "Manus's computer."

That transparency is one of the things that distinguished Manus from day one. Most AI agents are black boxes. Manus shows its work, and lets you intervene mid-task if it goes off track.

What happened in the last twelve months

The timeline is worth understanding because it moves fast and explains why Manus is suddenly everywhere.

March 2025: Public launch. Viral sensation. Invite-only, with a waitlist that hit 2 million. Benchmark scores that beat OpenAI's Deep Research on the GAIA evaluation. Partnership with Alibaba's Qwen team announced within a week.

April--November 2025: Steady product iteration. Task completion time dropped from over 15 minutes to under 4 minutes. Wide Research feature launched (parallel searches across multiple sources). Browser Operator for autonomous web navigation. Mobile app building from natural language. Revenue hit $100 million ARR by the end of this period -- genuine commercial traction, not just hype.

December 2025: Meta announced the acquisition. Reportedly $2 billion, though exact terms weren't disclosed. Meta said it would keep Manus running as a standalone product while integrating the technology into Meta AI. Manus founder Xiao Hong became a Meta VP.

January 2026: Chinese regulators announced a review of the deal, investigating whether Manus's technology falls under national security or technology export rules. The review is still pending.

February 2026: Two major moves. First, Manus launched personal AI agents in Telegram -- scan a QR code, connect your account, and get the full Manus platform inside a messaging app. WhatsApp, Slack, and Discord integrations were announced as coming within 30 days. Second, Meta quietly embedded Manus into Ads Manager, giving 4 million+ advertisers access to autonomous reporting, audience research, and campaign analysis.

March 2026: Manus has processed over 147 trillion tokens and created more than 80 million virtual computers since launch. It has 105 employees across Singapore, Tokyo, San Francisco, and is opening a Paris office. And the Chinese regulatory review remains unresolved.

What's actually working

Research and data compilation. This is where Manus consistently delivers. Give it a research brief -- competitive analysis, market sizing, vendor comparison -- and it produces structured, sourced output that would take a human analyst hours. The Wide Research feature runs parallel searches across multiple databases simultaneously, cross-references findings, and assembles everything into a coherent document. Users who compare it to ChatGPT's deep research mode generally find Manus faster and more thorough for multi-source tasks.

Background execution. Unlike most AI tools that require you to sit and watch, Manus works while you don't. Queue a task, close your laptop, come back later. This sounds minor but it fundamentally changes how you interact with an AI agent. Instead of a conversation, it's a delegation -- closer to assigning work to a contractor than chatting with an assistant.

The transparency model. The real-time replay of Manus's actions is genuinely useful. You can see exactly what it searched, which pages it opened, what data it extracted. This builds trust in a way that black-box agents can't, and it makes debugging straightforward when something goes wrong.

Messaging app integration. The Telegram launch in February was significant not for the technology but for the distribution model. Manus inside Telegram means an AI agent accessible to people who will never install a developer tool, never self-host anything, and never visit an AI company's website. If the WhatsApp integration ships as planned, Manus will be available inside the messaging app used by 2 billion people -- owned by the same company that now owns Manus.

What isn't working

The Ads Manager integration is rough. This is the most visible deployment of Manus, and early reports are not encouraging. Media buyers who tested it found outputs that hallucinated data -- fabricating metrics, misattributing campaign performance, generating analysis that sounded plausible but wasn't accurate. One agency VP told Digiday he's not sending any of the outputs to clients because they aren't reliable enough. For a tool embedded in a platform where wrong numbers can mean wrong spending decisions, this matters.

The regulatory overhang. Manus was built in China, headquartered in Singapore, and acquired by an American company. Chinese authorities are reviewing whether the deal complies with technology export and national security regulations. This review has been pending since January with no resolution. For enterprise users evaluating Manus, the uncertainty about whether the technology can continue operating in its current form is a real consideration.

Data flows through Meta. Once Manus became a Meta product, your data flows through Meta's infrastructure. For personal productivity tasks, this may not matter. For enterprise use cases involving sensitive data -- financial analysis, competitive intelligence, legal research -- the fact that Meta now sits between you and your AI agent's work product is a meaningful tradeoff. Particularly given Meta's history with data handling and the advertising business model that funds everything the company does.

Reliability at scale. Manus works well for tasks with clear structure: research briefs, data compilation, report generation. It struggles with tasks that require judgment, nuance, or domain expertise. The agent doesn't know your business context, your client relationships, or the political dynamics of your organization. It can produce a competitive analysis. It can't tell you which findings actually matter for your strategy. The gap between "completed the task" and "produced something useful" still requires a human to close.

The platform play

Here's what makes Manus strategically different from every other AI agent in the market: distribution.

OpenClaw has 235,000 GitHub stars but requires self-hosting. Claude Cowork requires a subscription and a desktop app. Perplexity Computer requires a $200/month Max plan. NanoClaw requires Claude Code fluency.

Manus is being pushed into platforms that billions of people already use. Ads Manager. Telegram. WhatsApp is next. Instagram and Facebook are likely after that. You don't need to know what an "AI agent" is. You don't need to install anything. It just appears in the tools you already have open.

This is the embedded-agent thesis playing out at a scale nobody else can match. Meta has 4 million advertisers in Ads Manager alone. WhatsApp has 2 billion users. If Manus ships reliably inside those platforms, the addressable market isn't "people who are interested in AI agents." It's "everyone."

That's a different game entirely from what OpenClaw, Cowork, and Perplexity Computer are playing.

Who Manus is actually for right now

Researchers and analysts who need to compile information from many sources into structured deliverables. This is where Manus has always been strongest.

Small businesses on Meta's platforms who want AI-assisted reporting and audience research without hiring an analyst. The Ads Manager integration is free to advertisers -- the value proposition is real, even if the reliability isn't there yet.

Non-technical users who want agent capabilities without setup overhead. The Telegram integration is the easiest on-ramp to an autonomous AI agent that currently exists.

Not yet: enterprise teams with strict data handling requirements. The Meta ownership, the Chinese regulatory uncertainty, and the hallucination issues in Ads Manager all create risk that most enterprise compliance teams won't accept today.

Not yet: anyone who needs accuracy guarantees. Manus completes tasks. It doesn't always complete them correctly. If you're in a domain where wrong outputs have real consequences -- finance, legal, healthcare -- you still need to verify everything Manus produces.

The bigger picture

A year ago, Manus was a demo that went viral. Today it's a $2 billion Meta product being embedded into platforms used by billions. The technology works well enough for research and data tasks. The Ads Manager rollout shows it's not ready for unsupervised commercial use. The regulatory situation adds uncertainty that could affect its future trajectory.

But the real story isn't about Manus specifically. It's about the distribution model. Every other AI agent -- OpenClaw, Claude Cowork, Perplexity Computer -- requires you to go somewhere and set something up. Manus is the first AI agent being pushed into platforms where people already are.

If Meta gets the reliability right, the implications are enormous. If it doesn't, the hallucination problem in a platform handling real advertising budgets could create real damage before anyone notices.

Either way, the AI agents that succeed in 2026 aren't the ones with the most impressive demos. They're the ones embedded in the workflows people already use. Manus is testing that thesis at a scale nobody else has attempted -- and the results so far are a mix of promising and premature.

This is part of a series on AI agents in 2026. See also: Perplexity Computer vs Claude Cowork, Perplexity Computer vs OpenClaw, Claude Cowork vs OpenClaw, and Best OpenClaw Alternatives That Don't Require Coding.