Manus AI Review: The Autonomous Agent That Seduced Meta for $2 Billion

Manus launched in March 2025 with a pitch that sounded too good to be true: an AI agent that doesn't just talk, but acts. Hand it a goal - research a competitor, build a website, analyze a dataset - and it goes off on its own, browsing the web, writing code, managing files, and delivering polished results. No hand-holding. No repeated prompts. Within weeks, it topped the GAIA benchmark, outscoring OpenAI's Deep Research. Within months, it claimed $100 million in annualized revenue. By December, Meta had bought the whole thing for over $2 billion. After spending several weeks testing Manus across research, coding, data analysis, and content tasks, I can say this: the highs are genuinely impressive. The lows are truly alarming. And the Meta acquisition changes everything about what this product is - and who it's really for.

What Manus Actually Is

If you've been tracking the AI agents space, you know that "agent" has become one of the most overloaded terms in tech. Chatbots, copilots, workflow automations, and genuine autonomous systems all crowd under the same umbrella. Manus sits firmly at the autonomous end.

Unlike a chat assistant that generates text in response to prompts, Manus operates in a cloud-based virtual computing environment with its own browser, shell, code interpreter, and file system. You describe a goal - "compile a competitive analysis of the top five CRM platforms," for instance - and Manus decomposes it into subtasks, delegates work to specialized sub-agents, browses the web, runs scripts, creates files, and delivers a finished output. Crucially, it runs asynchronously. You can close your laptop and come back later. The agent keeps working.

The architecture behind this is a multi-agent orchestration system. At the top sits an executor agent that manages the overall workflow. Below it, a planner agent breaks complex goals into manageable steps, and a knowledge agent handles information retrieval and synthesis. The system uses what the team calls a "CodeAct" approach - executable Python code as its primary action mechanism - which gives it more flexibility than agents limited to predefined tool calls.

Manus was built by Butterfly Effect, the Singapore-based (originally Beijing-based) company behind Monica.im. The engineering team has published some truly interesting technical writing about their approach to context engineering - treating the file system as extended memory, using KV-cache optimization as a first-class design concern, and deliberately leaving failed attempts in context so the model learns from its own mistakes within a session.

Manus delegates tasks to specialized sub-agents for browsing, coding, and analysis - coordinated by a central executor that manages the overall workflow.

Where Manus Delivers

The cases where Manus truly earns its keep cluster around structured, research-heavy tasks.

Deep research and analysis. This is Manus at its best. Ask it to compile a 200-page document summary, pull real-time market data, or cross-reference information across multiple sources, and it delivers results that are often more thorough than what you'd get from ChatGPT's Deep Research or Perplexity. In GAIA benchmark testing - a real-world problem-solving evaluation - Manus scored 86.5%, 70.1%, and 57.7% across three difficulty levels, consistently outperforming OpenAI's scores of 74.3%, 69.1%, and 47.6%. These aren't cherry-picked demos. The benchmark covers complex logic, multi-step reasoning, and dynamic decision-making.

In my own testing, I asked Manus to research the European AI Act's impact on open-source model licensing. It browsed regulatory databases, pulled recent enforcement actions, cross-referenced academic papers, and produced a structured report with citations. The whole thing ran in the background while I was at lunch. The output wasn't perfect - some secondary sources were outdated - but the breadth and structure of the research would have taken me several hours to assemble manually.

Data analysis and file creation. Manus can download datasets, run Python analysis scripts, create visualizations, and package everything into downloadable reports. This is a meaningful differentiator from chat-based tools. I gave it a CSV of 50,000 customer support tickets and asked for a sentiment analysis with trend visualization. It installed the necessary Python libraries, wrote analysis code, generated charts, and delivered a formatted PDF - all without intervention.

Multi-step web tasks. Need to check pricing across a dozen SaaS competitors? Monitor job postings from specific companies? Compile a list of investors who've funded AI startups in the last six months? These are the kinds of tasks where Manus's autonomous browsing and data extraction shine. It doesn't just search - it navigates, clicks, scrolls, and extracts structured data from live websites.

Where Manus Falls Short

The gap between Manus's best work and its worst is wider than the Strait of Messina.

Reliability is the biggest problem. The most consistent complaint across user reviews, forums, and my own testing is that Manus simply isn't stable enough for professional use. I hit the "Due to current high service load, tasks cannot be created" wall multiple times during peak hours. Sessions crashed. Long-running tasks died mid-execution with no way to recover partial results. One multi-step research project that was two hours into execution just vanished. Trustpilot reviews tell similar stories - users report paying for credits, only to watch them evaporate on tasks that loop endlessly or crash without producing output.

Credit consumption is a black box. More on pricing below, but the core issue is that you have no way to estimate what a task will cost before you start it. A simple research query might burn 50 credits. A complex data analysis task could consume 900. There's no preview, no budget cap, no "this will cost approximately X" warning. For a product that charges based on consumption, this opacity is a serious problem.

Coding output is inconsistent. While Manus can write functional code, it's not competitive with dedicated coding agents like Devin or IDE-integrated tools. The code it produces works for prototyping and one-off scripts, but production-quality engineering tasks often come back with questionable architecture decisions, missing error handling, or outright bugs. If coding is your primary use case, you'll get better results from purpose-built tools.

Customer support is nearly nonexistent. Multiple users report automated responses, copy-paste replies, and an inability to reach a human agent for billing disputes. Several Trustpilot reviewers describe being charged hundreds of dollars without authorization and finding no path to resolution. For a product backed by a $2 billion acquisition, the support infrastructure feels like it belongs to a beta startup.

Pricing - The Credit Labyrinth

Manus operates on a credit-based system that's both flexible and deeply frustrating.

Manus's credit-based pricing means costs vary wildly depending on task complexity - with no way to preview what a task will cost before starting it.

The current plans:

• Free: Limited monthly credits, access to core features
• Standard ($20/month): 4,000 credits, 300 daily refresh credits, up to 20 concurrent tasks
• Customizable ($40/month): 8,000 credits, same daily refresh and concurrency limits
• Extended ($200/month): 40,000 credits for heavy users

Annual billing gets you a 17% discount. All plans include 300 credits that refresh daily, which is a nice touch for light users.

The problem isn't the pricing tiers themselves - they're reasonable compared to other agent platforms. The problem is unpredictability. A single complex task can burn 500-900 credits. Credits don't roll over. And because there's no cost preview, budgeting is effectively guesswork. I spoke to three independent consultants who use Manus regularly, and none of them could predict their monthly bill within a 50% margin.

Compare this to a flat-rate subscription like Perplexity Pro at $20/month for unlimited queries, and the value proposition gets murky fast. Manus can do more than Perplexity - it can execute code, create files, and perform multi-step workflows. But if you can't predict the cost, "more capability" doesn't automatically mean "better value."

The Meta Acquisition - What Changes

On December 29, 2025, Meta bought Manus for over $2 billion. To put that in perspective: Manus had a $500 million valuation just eight months earlier, when Benchmark led a $75 million funding round in April 2025. A 4x return in under a year.

Meta's $2 billion acquisition of Manus - its fifth AI-focused purchase of 2025 - signals a major bet on autonomous agents as the next interface layer for its platforms.

The strategic logic is clear. Meta wants AI agents that can act, not just chat. Manus has already processed over 147 trillion tokens and spun up 80 million virtual computing sessions. Meta wants to plug that execution capability into its ecosystem - Facebook, Instagram, WhatsApp, and especially its advertising platform. In fact, Meta has already started rolling out Manus within Ads Manager for some advertisers, automating tasks like campaign reporting and market research.

For existing Manus users, the company has said the product will continue operating as a standalone service. The Singapore-based team remains in charge of day-to-day operations, and CEO Xiao Hong has framed the deal as building "on a stronger, more sustainable foundation."

But here's the thing about acquisitions: the product always changes. Meta didn't spend $2 billion to maintain the status quo. The new Meta Superintelligence Labs (MSL) division, where the Manus team will lead development of agentic features across the Meta suite, makes the direction clear. Manus's best engineering talent will increasingly be pulled toward Meta's priorities - advertising automation, business messaging, and enterprise tooling. Whether the standalone product gets the same love remains to be seen.

The acquisition also comes with geopolitical baggage. Manus was originally founded in Beijing before relocating to Singapore. China has opened a formal investigation into the deal, looking at potential violations of tech export controls, cross-border data flows, and national security concerns. Beijing isn't happy about Chinese-origin AI technology ending up under American corporate control. The deal's completion may depend on regulatory outcomes that have nothing to do with the product's quality.

How It Compares

In the rapidly growing AI agent market, Manus occupies a distinct niche: the generalist autonomous agent.

Manus vs. Devin: Devin is specialized for software engineering - it writes, debugs, tests, and launches code in a sandboxed environment. Manus is broader but shallower on coding. If your primary need is code, Devin wins. If you need research, analysis, web scraping, and content generation in addition to occasional coding, Manus covers more ground. Devin also starts at $20/month with predictable ACU-based billing, while Manus's credit system is harder to forecast.

Manus vs. ChatGPT Deep Research: ChatGPT produces excellent research reports but can't execute code, create downloadable files, or perform multi-step web automations. Manus can. The tradeoff is reliability - ChatGPT rarely crashes mid-task, while Manus does regularly.

Manus vs. Perplexity: Perplexity excels at research with clean source attribution. It's faster and more reliable for pure information gathering. Manus goes further by actually executing on what it finds - creating files, running analysis, building prototypes. Different tools for different jobs.

Strengths

• Deep research quality is truly best-in-class, outperforming OpenAI on GAIA benchmarks across all difficulty levels
• Asynchronous execution means you can delegate tasks and walk away - a fundamentally different workflow from chat-based tools
• Multi-modal task breadth spanning research, coding, data analysis, web scraping, and content creation in a single platform
• Cloud-based sandboxing gives each session a full Linux environment with browser, shell, and Python interpreter
• Technical architecture is thoughtful - the context engineering, KV-cache optimization, and error retention approaches are genuinely novel
• Cross-platform availability with web, iOS, Windows, browser extension, Slack, and Telegram integrations

Weaknesses

• Server reliability is unacceptable for a paid product - frequent overload errors, session crashes, and lost work
• Credit-based pricing with no cost preview makes budgeting impossible and erodes trust
• Customer support ranges from unhelpful to nonexistent, with billing disputes going unresolved
• Code quality is inconsistent and not competitive with dedicated coding agents
• Post-acquisition uncertainty raises questions about the standalone product's future under Meta
• Geopolitical risk from the ongoing Chinese regulatory investigation could affect operations

The Verdict - 6/10

Manus is the most ambitious general-purpose autonomous agent on the market, and on its best day, it delivers results that no competitor can match. The deep research capability is legitimately impressive. The ability to delegate complex, multi-step tasks and come back to finished results is a glimpse of what AI agents are supposed to be. The technical architecture - especially the context engineering and multi-agent orchestration - shows genuine engineering sophistication.

But a tool is only as good as its worst day, and Manus's worst days are bad. Crashed sessions, eaten credits, vanishing work, and a customer support experience that feels like shouting into a void. The credit-based pricing model punishes exactly the kind of complex tasks where Manus is supposed to excel. And the Meta acquisition, while confirming the technology, introduces uncertainty about the product's long-term direction as an independent service.

At $20/month for the Standard plan, Manus is worth experimenting with - especially if your work involves regular research, data analysis, or structured information compilation. The free tier lets you test the waters before committing. But if you need a tool you can depend on for daily professional work, the reliability and support gaps are hard to overlook.

The underlying technology is excellent. The execution - from infrastructure to billing to support - needs to catch up. Meta has the resources to fix these problems. Whether they choose to invest in the standalone product or redirect that talent toward advertising automation will determine whether Manus becomes the general-purpose AI agent it promises to be, or just another feature inside the Meta ecosystem.

Sources

• Manus AI Review 2026 - CyberNews
• Meta acquires Manus AI - TechCrunch
• Meta's $2B Manus deal pushes away customers - CNBC
• Context Engineering for AI Agents - Manus Blog
• Manus Joins Meta - Manus Blog
• China deepens review of Meta's Manus buyout - Bloomberg
• GAIA Benchmark analysis - Helicone
• Why Meta bought Manus - VentureBeat
• Manus AI technical investigation - GitHub
• Meta rolling out Manus in Ads Manager - The Keyword