Key Takeaways
- Manus AI is a general-purpose autonomous AI agent built by Butterfly Effect, a startup founded by Xiao Hong, and launched on March 6, 2025 in invitation-only beta.
- The agent went viral within 72 hours of launch, with over 2 million people joining the waitlist and invite codes reportedly reselling for up to 50,000 CNY (roughly $7,000 USD) on secondary markets.
- Manus reported self-assessed GAIA benchmark scores of 86.5% at Level 1 and 57.7% at Level 3, surpassing OpenAI Deep Research (74.3% and 47.6% respectively), though these scores have not been independently verified.
- Meta announced an acquisition of Manus in December 2025 for roughly $2 billion. China’s National Development and Reform Commission blocked the deal in April 2026, ordering the parties to unwind the transaction. The acquisition never completed.
- Manus uses a credit-based pricing model: Free (300 daily refresh credits), Pro at $20/month (4,000 credits), and higher tiers. Credits do not roll over and complex tasks can deplete a monthly allocation fast.
- Manus orchestrates multiple AI models, including fine-tuned versions of Alibaba’s Qwen and Anthropic’s Claude 3.5 Sonnet, rather than relying on a single LLM, which is what separates it architecturally from ChatGPT and Claude.
- Common user complaints include unpredictable credit burn, frequent crashes under high server load, getting stuck on paywalled sites and CAPTCHAs, and limited integrations with third-party business tools.
- Butterfly Effect relocated its headquarters from China to Singapore in mid-2025, retaining about 40 core technical staff out of roughly 120 employees, after Meta’s acquisition plans emerged.
- Manus opened general access to all users (no invite required) in May 2025, ending the initial waitlist period.
When Manus AI dropped its first demo video on March 6, 2025, showing an autonomous agent screening job applications and doing stock research without a single human prompt, the internet broke. Within 72 hours, the waitlist had crossed six figures. Within weeks, it had passed 2 million. People were paying scalpers thousands of dollars for an invite code. That kind of reception had not happened in the AI space since ChatGPT itself.
The buzz was partly justified and partly hype. Manus AI genuinely does something different from standard chatbots: it takes a high-level goal and works through it step by step, browsing the web, writing and running code, creating files, and delivering a finished output without you having to babysit the process. But the real-world experience, as dozens of testers found out, is messier than the polished demo.
This review covers what Manus actually is, what it can and cannot do, how pricing works, how it compares to GPT-4o and Claude, and the full story of Meta’s attempted $2 billion acquisition that was ultimately blocked by China. If you are deciding whether to subscribe, or just trying to understand what the fuss was about, this is the rundown you need.
What is Manus AI?
Manus AI is a product of Butterfly Effect, a startup founded by Xiao Hong in 2022, two months before OpenAI publicly launched ChatGPT. Xiao, originally from Jian in Jiangxi province, studied at Huazhong University of Science and Technology in Wuhan, where he led student teams building in-campus tools before graduating into entrepreneurship full-time.
Butterfly Effect’s first notable product was Monica, a browser extension released in 2023 that aggregated multiple commercial language models into a single interface for translation, summarization, and writing. Monica crossed one million users quickly, mostly in overseas markets, and gave the team the revenue and credibility to fund something bigger.
That something bigger was Manus. The name comes from the Latin word for “hand,” chosen deliberately to contrast with the concept of “mind.” The idea is that an AI can be extremely capable as a reasoning engine (the mind) but remain useless unless it can also act in the world (the hand). Manus was built to close that gap.
Unlike a chatbot that gives you an answer you must then act on, Manus receives a task, plans a series of steps, executes those steps using a suite of tools (browser, code interpreter, file system, external APIs), and returns a completed deliverable. It is closer to hiring a contractor than querying a search engine.
After raising a Series B funding round that valued the company at $500 million, Butterfly Effect relocated its headquarters to Singapore in mid-2025. Around 40 core technical staff moved with the company; the remaining roughly 80 China-based employees were let go as the company pivoted to international operations.
Manus AI Features
Multi-Agent Architecture
Manus does not run on a single language model. It orchestrates multiple specialized AI models, including fine-tuned versions of Alibaba’s open-source Qwen models and Anthropic’s Claude 3.5 Sonnet, assigning different sub-tasks to whichever model is best suited for the job. An orchestrator agent breaks a user’s goal into steps, delegates each step to sub-agents, and assembles the outputs into a final result. This is architecturally more complex than a single-model assistant, which is why Manus can tackle longer, multi-step workflows.
Autonomous Web Browsing
Manus can navigate websites, fill out forms, extract data from pages, and follow links across multiple sessions without human guidance. In practical terms, this means you can ask it to compile a competitive analysis of five SaaS companies and it will actually visit each company’s pricing page, pull the numbers, and format them into a report. It does not just describe how you might do this research; it does the research.
Code Writing and Execution
The agent can write code in Python, JavaScript, and other languages, execute it in a sandboxed environment, review the output, and iterate on errors automatically. Users have successfully used this for data cleaning scripts, web scrapers, basic API integrations, and exploratory data analysis. The limitation here is on complexity: Manus handles well-defined coding tasks but struggles with large, multi-file software projects.
File Creation and Management
Manus can create, edit, and organize files, including spreadsheets, documents, PDFs, and structured data files. A common use case is asking it to research a topic, compile findings into a formatted report, and deliver a downloadable file, all in one prompt.
Memory and Context Retention
Manus retains key instructions as “knowledge” across sessions, so you do not have to re-explain your preferences or context every time. This is more useful in ongoing workflows than in one-off tasks and is one of the features that distinguishes Manus from simple API wrappers.
Task Transparency
As Manus works, it shows you what it is doing in real time: which sites it is visiting, what code it is writing, and what decisions it is making. Users familiar with black-box AI tools have noted that this transparency makes it easier to trust the process and catch errors early.
Manus AI Benchmark Performance
Manus reported GAIA benchmark results when it launched: 86.5% at Level 1, 70.1% at Level 2, and 57.7% at Level 3, according to Helicone’s benchmark analysis. For context, OpenAI’s Deep Research scored 74.3%, 69.1%, and 47.6% on the same three levels.
These figures are compelling if taken at face value. The important caveat is that these are self-reported scores and have not been independently re-verified by third parties since the initial launch. Benchmark inflation is a known issue in the AI industry, and Manus has not published full methodology documentation that would allow external researchers to replicate the results. Treat the numbers as directionally interesting, not as certified fact.
Manus AI Pricing
Manus uses a credit-based system, which is one of the most discussed and debated aspects of the product. According to Manus’s official pricing page and verified by eesel AI’s pricing analysis:
- Free Plan: 300 daily refresh credits that reset each day. Access to Manus 1.6 Lite in Agent Mode. Good for light, exploratory use.
- Pro ($20/month): 4,000 monthly credits plus the 300 daily refresh credits. Annual billing saves approximately 17%.
- Pro ($40/month): 8,000 monthly credits plus daily refresh credits. Better for heavier recurring use.
- Team Plan: Starting at $20 per seat per month with a shared team credit pool and admin controls.
The credit system has attracted significant criticism. A single moderately complex research task can consume hundreds of credits, and users on the $40/month plan report their monthly allocation disappearing within a week of heavy use. Monthly credits do not roll over, which means any unused credits at the end of the billing period are lost. The Pro plan at $199/month provides 19,900 credits, but users report that complex tasks burn through 900 credits or more in a single session.
The Meta Acquisition Story
In December 2025, Meta announced it would acquire Manus for an amount reported by multiple sources to be over $2 billion, making it one of the largest AI acquisitions Meta had attempted. According to TechCrunch’s reporting, Meta’s stated goal was to accelerate AI innovation for businesses and integrate autonomous agent capabilities into its consumer and enterprise products, including Meta AI.
Manus had by that point reached an annualized revenue run rate exceeding $125 million, just eight months after launch. The deal appeared significant on both sides: Meta needed agentic AI capabilities to compete with OpenAI and Google, and Manus needed the infrastructure and distribution that Meta could provide.
The deal never closed. On April 27, 2026, China’s National Development and Reform Commission (NDRC) blocked the acquisition, ordering both parties to unwind the transaction, as reported by TechCrunch and CNBC. The NDRC’s position was that relocating corporate registration to Singapore did not place a Chinese-founded company beyond China’s regulatory reach, particularly when the underlying technology and research ecosystem remained tied to the mainland. The article title references “Meta acquired” Manus: to be accurate, Meta announced the acquisition and it was widely reported, but the deal was subsequently blocked. As of May 2026, the acquisition has not been completed.
Manus continues to operate as an independent company, running its subscription service without interruption.
Manus AI Pros and Cons
Pros:
- Genuinely autonomous task completion: it plans, executes, and delivers without constant prompting
- Multi-model architecture produces better results on diverse tasks than any single LLM alone
- Transparent operation: you can watch exactly what the agent is doing at each step
- Handles real research workflows that pure chatbots cannot, including live web browsing and file creation
- Memory retention across sessions reduces repetitive setup for recurring workflows
- GAIA benchmark results, while self-reported, suggest strong autonomous task performance
- Free tier is generous enough to evaluate the product before committing to a paid plan
Cons:
- Credit system is expensive relative to output, especially for complex tasks
- Struggles with paywalled content and CAPTCHA challenges, limiting research depth
- Crashes and “high service load” errors are common, particularly during peak usage windows
- Not suitable for nuanced creative writing, sensitive communication, or complex software development
- Monthly credits do not roll over, penalizing inconsistent users
- Limited integrations with third-party business tools compared to more mature automation platforms
- GAIA benchmark scores are self-reported and unverified independently
- The future of the company remains uncertain following the failed Meta acquisition
Manus AI vs Alternatives
Manus sits in a category with a handful of other agentic AI tools, each with a different approach.
Manus vs ChatGPT (GPT-4o): ChatGPT is a conversational assistant that can use tools when prompted but does not autonomously plan and execute multi-step workflows the way Manus does. For a single question, a writing task, or a coding problem, ChatGPT is faster and cheaper. For a research-heavy deliverable requiring web browsing, data extraction, and file creation, Manus has the edge. Manus’s own comparison page lays this out, though obviously with a bias toward their product.
Manus vs Claude: Claude 3.5 Sonnet, ironically, powers part of Manus’s own architecture. Claude excels at long-context reasoning, careful instruction-following, and nuanced writing, with a 200,000-token context window. Claude does not execute autonomous multi-step tasks the way Manus does, though Anthropic has expanded Claude’s tool-use capabilities significantly since Manus launched.
Manus vs OpenAI Deep Research: OpenAI’s Deep Research feature is the closest direct competitor, focused specifically on research tasks with web browsing. Manus’s self-reported GAIA scores beat Deep Research, but Deep Research is bundled into ChatGPT Pro ($20/month) without a separate credit cost per task, which many users find more predictable for budget planning.
Manus vs Google Gemini: Gemini 2.0 Flash has strong multimodal capabilities and is tightly integrated into the Google ecosystem. For users who live in Google Workspace, Gemini offers deeper native integrations. Manus offers more autonomous action-taking but less ecosystem integration.
Who is Manus AI Best For?
Manus works well for a specific type of user. If you regularly need research compiled into formatted deliverables, competitive analyses pulled from live web data, or data-processing scripts written and run without your involvement, Manus can save significant time. Consultants, analysts, and founders who do repetitive research-heavy workflows are the clearest beneficiaries.
It is not the right tool if your needs are primarily creative writing, sensitive email drafting, complex multi-file software development, or tasks that require integrating with tools like Salesforce, Notion, or Slack out of the box. For those workflows, purpose-built tools or a well-prompted Claude or ChatGPT will serve you better at lower cost.
Budget-conscious users should be cautious. The free tier is a real test-drive, but the paid plans require discipline about task complexity if you want your monthly credits to last. Power users running multiple complex research tasks per day will find even the $40/month Pro plan insufficient.
Our Verdict
Manus AI delivered something genuine when it launched: an autonomous agent that actually completes tasks rather than describing how you might complete them yourself. For the narrow category of research-heavy, multi-step work that chatbots handle poorly, Manus is meaningfully better than the alternatives as of mid-2025.
The product has real rough edges. Credit costs are opaque and can escalate fast. Server reliability under load is inconsistent. The tool gets stuck on paywalls and CAPTCHAs with surprising frequency for a product positioning itself as a research agent. And the uncertainty around the failed Meta acquisition raises questions about Manus’s long-term roadmap.
That said, the free tier is substantial enough to test whether Manus fits your workflow before spending a dollar. For users who find it useful, the $20/month Pro entry point is reasonable if your tasks are not too credit-intensive. For heavy users, the math gets harder to justify against alternatives like ChatGPT Pro, which bundles Deep Research without per-task credit costs.
Rating: 3.8 out of 5. Genuinely useful for a specific user profile, but overhyped for general-purpose use, and the credit system needs rethinking before this can be recommended without reservation.
Frequently Asked Questions
What is Manus AI and how is it different from ChatGPT?
Manus AI is an autonomous agent that plans and executes multi-step tasks independently, including browsing the web, writing and running code, and creating files, all without continuous human prompting. ChatGPT is primarily a conversational assistant that answers questions and uses tools when asked, but does not autonomously chain together dozens of actions to complete a long-horizon task. The core difference is agency: Manus does the work, while ChatGPT tells you how the work could be done.
Did Meta actually acquire Manus AI?
Meta announced an acquisition of Manus in December 2025, with the deal reported at roughly $2 billion by sources including TechCrunch and Bloomberg. However, China’s National Development and Reform Commission blocked the deal in April 2026, ordering both parties to unwind the transaction. As of May 2026, the acquisition has not been completed and Manus continues to operate as an independent company.
How much does Manus AI cost?
Manus offers a free plan with 300 daily refresh credits. Paid plans start at $20/month for 4,000 monthly credits and $40/month for 8,000 monthly credits. A Team plan starts at $20 per seat per month. Annual billing saves approximately 17%. Credits do not roll over at the end of each billing period.
Is Manus AI still invite-only?
Manus launched in invite-only beta on March 6, 2025, with more than 2 million people joining the waitlist in the first few weeks. General access opened in May 2025, allowing anyone to sign up using email, Google, Apple, or Microsoft accounts without requiring an invitation.
What tasks is Manus AI best at?
Manus performs best on research-heavy, multi-step tasks that require browsing multiple websites, extracting data, running code, and producing a formatted deliverable. Common strong use cases include competitive analysis reports, financial research summaries, data scraping and cleaning, and automated content compilation. It is not well suited to creative writing, complex software development, or workflows requiring deep integration with business tools like Salesforce or Notion.
What are the main complaints about Manus AI from real users?
The most frequent complaints are about credit consumption (complex tasks drain credits faster than expected), server instability (the “high service load” error is common during busy periods), and the agent getting stuck on paywalled sites and CAPTCHAs. Some users also report agents looping on a single page rather than progressing, and context length limits forcing them to break large projects into smaller chunks manually.
How does Manus AI compare to OpenAI’s Deep Research?
Manus reported GAIA benchmark scores of 86.5% at Level 1 and 57.7% at Level 3, versus OpenAI Deep Research at 74.3% and 47.6%, though these are self-reported figures from Manus and have not been independently verified. In practical terms, both tools are designed for research workflows requiring web browsing. OpenAI Deep Research is bundled into ChatGPT Pro at $20/month without per-task credit costs, which many users find more budget-predictable than Manus’s credit system.
Who built Manus AI?
Manus was built by Butterfly Effect, a startup founded by Xiao Hong in 2022 and originally based in Wuhan and Beijing, China. The company also operates Monica, a browser extension that aggregates multiple AI models and reached over one million users before Manus launched. After raising a Series B at a $500 million valuation, Butterfly Effect relocated its headquarters to Singapore in mid-2025.
Is Manus AI safe to use for sensitive data?
Manus processes tasks in a cloud environment, which means data you provide or that the agent collects during research passes through Butterfly Effect’s servers. The company has published a privacy policy, but users handling sensitive business data, confidential documents, or personally identifiable information should review the terms carefully before using Manus for those workflows. No autonomous cloud agent should be used with data you would not share with a third-party SaaS provider.
What happened to Manus AI after the Meta deal fell through?
After China’s NDRC blocked the Meta acquisition in April 2026, Manus continued operating its subscription service without interruption. The company has not made public statements about alternative acquisition plans or new fundraising rounds. The product roadmap and team composition remain the same as before the acquisition announcement, and users continue to access the service normally on their existing plans.
Manus AI is a legitimate step forward in autonomous AI agents, not just another chatbot dressed up with a fancy interface. For researchers, analysts, and anyone whose work involves pulling together information from across the web and packaging it into usable outputs, it is worth at least a free-tier test. The limitations are real and the credit system needs work, but the core capability is genuine.
If you want to try it, start with the free plan at manus.im and run a few real tasks from your actual workflow before committing to a paid tier. That will tell you more than any review can.
