Manus (AI agent)
Last edited
Fact-checked
In review queue
Sources
30 citations
Revision
v6 · 3,450 words
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Manus is a general-purpose autonomous AI agent developed by Butterfly Effect, a startup founded in China in 2022 and later headquartered in Singapore. Launched in an invitation-only preview on March 6, 2025, Manus is marketed as a "general AI agent" that executes multi-step tasks asynchronously in the cloud rather than answering single prompts, and it drew intense global attention within days of its debut. The product is built as an orchestration layer on top of existing frontier models, principally Anthropic's Claude and fine-tuned versions of Alibaba's open-source Qwen, rather than a wholly in-house foundation model. Butterfly Effect is also the company behind the Monica AI browser-assistant, and its founders include chief executive Xiao Hong and chief scientist Ji Yichao (known as "Peak").[1][2][3][4]
In April 2025 Butterfly Effect raised roughly US$75 million in a round led by the venture firm Benchmark, valuing the company at about US$500 million. Manus then scaled quickly: the company said it reached US$100 million in annual recurring revenue about eight months after launch, which it called being "the fastest startup to go from $0 to $100M in the world."[25] That investment, made into a Chinese-founded startup, later drew United States regulatory scrutiny, and on December 30, 2025 Meta agreed to acquire Butterfly Effect in a deal reported at around US$2 billion. China's economic planning agency, the National Development and Reform Commission, moved to block the acquisition on April 27, 2026, and the transaction was formally unwound in June 2026, making Manus a focal point in the wider contest over cross-border artificial-intelligence investment.[5][6][7][8][25][26]
History
Company and founders
Butterfly Effect was founded in 2022 by Xiao Hong, a Chinese entrepreneur and software engineer who studied software engineering at Huazhong University of Science and Technology in Wuhan. The company's first product, launched in 2023, was Monica, a browser extension that placed several commercial large language models behind a single interface for translation, summarization, and writing assistance. Monica was aimed primarily at users outside China and gave the company experience building consumer products on top of third-party models, an approach it carried over into Manus.[1][2][9]
The company's leadership combined product and research backgrounds. Xiao Hong served as chief executive and set strategy and product direction. Ji Yichao, who goes by "Peak," joined as co-founder and chief scientist; born in 1992 and raised partly in Colorado and Beijing, he had earlier founded Peak Labs (backed by ZhenFund and Sequoia China) and built the Magi search engine, and was named to Forbes China's 30 Under 35 lists. Zhang Tao, who had held product roles at companies including Tencent and ByteDance, led product.[1][2][9]
| Person | Role | Background |
|---|---|---|
| Xiao Hong | Founder and chief executive | Software engineering, Huazhong University of Science and Technology; founder of Monica (2022) |
| Ji Yichao ("Peak") | Co-founder and chief scientist | Founder of Peak Labs and the Magi search engine; Forbes China 30 Under 35 |
| Zhang Tao | Product lead | Former product roles at Tencent, ByteDance, and others |
Launch and virality
Manus launched in an invitation-only beta on March 6, 2025. A demonstration video showing the agent autonomously completing tasks such as resume screening and stock analysis drew more than one million views within about twenty hours, and the closed beta's waitlist climbed into the hundreds of thousands within days; the company has said more than two million people registered for access. A Butterfly Effect representative publicly conceded that the team had "completely underestimated" the level of public enthusiasm, noting the launch had carried essentially no marketing budget.[2][3][10]
The scarcity of invitation codes produced a brisk secondary market. Codes were resold on Chinese platforms such as Xianyu and on overseas resale sites, with reported prices ranging from a few thousand US dollars to, by some accounts, tens of thousands of yuan (figures cited in the range of ¥50,000 to ¥100,000, roughly US$7,000 to US$13,800 at the time). The Wall Street Journal reported that codes changed hands for more than US$1,000. The frenzy prompted widespread comparisons to DeepSeek, the Chinese model lab whose low-cost releases had caused a market shock weeks earlier.[3][10][11]
Funding
Before its breakout round, Butterfly Effect had raised earlier capital from Chinese investors. Reporting describes seed and early-stage backing from ZhenFund, and investment from HongShan (the firm formerly known as Sequoia China) and Tencent, with cumulative early funding described as somewhat north of US$10 million. According to 36Kr, the company turned down a roughly US$30 million acquisition approach from ByteDance in 2024, and its founders later declined offers of investment from several Chinese local governments out of concern that domestic state ties could invite scrutiny in Western markets.[1][5][6]
In April 2025, on the strength of the Manus launch, the company raised approximately US$75 million in a round led by Benchmark, valuing Butterfly Effect at about US$500 million, roughly a fivefold increase over its prior valuation. Bloomberg reported that the company planned to use the proceeds to expand into new markets including the United States, Japan, and the Middle East. Participants reported across sources included existing backers; some accounts also named Meituan co-founder Wang Huiwen among investors.[5][6]
| Round | Date | Amount | Lead / notable investors | Valuation |
|---|---|---|---|---|
| Early-stage (seed / Series A) | 2023-2024 | "north of" US$10M (cumulative, reported) | ZhenFund, HongShan (formerly Sequoia China), Tencent | Not disclosed |
| Series B | April 2025 | ~US$75M | Benchmark | ~US$500M |
Sources differ on the precise structure and timing of the earlier rounds. Some accounts place a roughly US$10 million Series A in 2023, while others describe a seed round in 2023 followed by a later Series A; the cumulative figure and the principal early investors (ZhenFund, HongShan, Tencent) are consistently reported.[1][5][6]
Relocation to Singapore and layoffs
In mid-2025, following the Benchmark round, Butterfly Effect shifted its headquarters to Singapore and moved its international operating entity to a Singapore-registered company (Butterfly Effect Pte. Ltd.). As part of the restructuring, the company cut its mainland China team from roughly 120 staff to about 40 core technical employees, who were relocated to Singapore, while the remaining staff were laid off with reported severance packages. Manus also closed its Chinese-language social-media accounts, blocked mainland China IP addresses from its website, and shelved a planned China-specific version of the product.[12][13][14]
Commentators described the move as an instance of "Singapore washing," a label for Chinese-linked companies that reincorporate or relocate abroad to ease fundraising and reduce exposure to geopolitical risk. The relocation was widely read as a response to tightening United States rules on outbound investment in Chinese AI, semiconductor, and quantum-computing firms, and to the regulatory questions surrounding the Benchmark investment.[12][13][14]
How does Manus work?
Manus runs as an autonomous agent in a cloud sandbox. When given a task, it spins up a virtual machine equipped with tools such as a web browser, a code interpreter, office applications, and design engines, then plans and carries out the work over many steps without continuous human prompting. A panel sometimes called "Manus's Computer" lets users watch the agent's actions in real time and intervene if needed, and because tasks run asynchronously in the cloud, a user can close the session and return to a completed result later.[4][15]
Technically, Manus is an orchestration layer rather than a new foundation model. Multiple independent reports, and the company's own statements, describe it as combining existing and fine-tuned models, principally Anthropic's Claude (reported as Claude 3.5 Sonnet at launch) together with fine-tuned versions of Alibaba's open-source Qwen, coordinated by Manus's own planning, tool-use, and context-management software. Analysts have characterized the system's distinctive engineering as "context engineering," the management of what information the underlying models see at each step, layered over the third-party models. Because it depended on American models not readily available in China, Manus was designed from the outset for markets outside the mainland.[1][3][4][15]
This architecture drew a pointed contrast with DeepSeek. Whereas DeepSeek had trained and openly released its own models, Butterfly Effect did not develop foundation models in-house and did not open-source the underlying technology, which led many observers, including TechCrunch, to argue that Manus was not a comparable "DeepSeek moment" but rather a sophisticated wrapper and product layer over existing systems.[3][11]
How does Manus perform on benchmarks?
At launch the company said Manus set state-of-the-art results on GAIA, a benchmark for general AI assistants developed by researchers at Meta AI, Hugging Face, and others that tests an agent's ability to reason, browse the web, and use tools on real-world tasks. Chief scientist Ji Yichao stated that Manus outperformed OpenAI's Deep Research on GAIA, and company-reported figures cited scores around 86.5% on the benchmark's easier (Level 1) tier and roughly 70% on a harder tier, ahead of the prior published state of the art. These results were self-reported and were not independently audited, a caveat noted by multiple outlets.[3][10][16]
How has Manus evolved since launch?
Manus shipped a rapid series of upgrades through 2025. In July 2025 it introduced Wide Research, a mode that deploys many parallel sub-agents (reported at more than 100), each a general-purpose Manus instance, to split a large job into pieces and work on them simultaneously, for tasks such as comparing dozens of products or generating many design variations at once. Wide Research launched first on the top Pro tier before extending to lower-priced plans.[27]
On October 16, 2025 the company released Manus 1.5, which it described as a re-architecture focused on speed and reliability. The company said tasks completed "nearly four times faster," dropping from about 15 minutes in April to under 4 minutes, alongside a roughly 15% gain in task quality and an expanded context window for single tasks. Manus 1.5 also added a full-stack application builder able to produce production-ready web apps with a backend, user authentication, a database, notifications, and analytics from a single prompt, and it split into two agents: Manus-1.5, the full-power model, and Manus-1.5-Lite, a lower-cost variant that consumes credits at a 50% discount.[23]
On December 15, 2025 Manus released version 1.6, led by a new flagship agent called Manus 1.6 Max. The company reported that user satisfaction rose by more than 19% in double-blind testing and that more tasks now completed autonomously without human intervention. The update added end-to-end mobile app development and Design View, an interactive canvas for point-and-click image editing and composition, across macOS, Windows, iOS, and Android.[24]
| Release | Date | Highlights |
|---|---|---|
| Public beta | March 6, 2025 | Invitation-only autonomous agent |
| Paid launch and iOS app | March 31, 2025 | Credit-based subscriptions |
| Wide Research | July 2025 | Many parallel sub-agents for large-scale tasks |
| Manus 1.5 / 1.5-Lite | October 16, 2025 | ~4x faster tasks, expanded context, full-stack app builder |
| Manus 1.6 / 1.6 Max | December 15, 2025 | Higher autonomy, mobile app development, Design View |
Reception
Early hands-on coverage was a mix of genuine intrigue and skepticism. In a widely cited test, MIT Technology Review found Manus highly intuitive and useful for bounded analytical research, comparing its output to "the sorts of things a skilled human intern could do during a day of work," and noted a cost advantage of roughly US$2 per task against about US$20 for some competitors. The same review, however, reported frequent crashes, server-overload messages, trouble with captchas and paywalls, and a higher failure rate than OpenAI's ChatGPT Deep Research.[4]
TechCrunch's testing was harsher, documenting crashes on simple tasks (such as ordering food), failed flight and restaurant bookings that returned broken links, factual errors, and inconsistent citation of sources; a company spokesperson responded that the closed beta's main purpose was to "stress-test various parts of the system and identify issues." Manus was frequently compared with agentic competitors such as OpenAI's Operator and Deep Research and Google's Project Mariner, with proponents emphasizing that Manus aimed to run tasks fully autonomously rather than pausing for user approval at each step.[3][11]
On March 11, 2025, Manus announced a strategic partnership with Alibaba's Qwen team, saying the two sides would collaborate on Qwen's open-source models to bring the agent's capabilities to Chinese models and computing platforms, a deal seen as helping Manus manage its traffic surge while strengthening Alibaba against rivals. The China-specific version tied to that partnership was later shelved amid the company's relocation.[15][17]
How much does Manus cost?
Manus moved out of closed beta into paid subscriptions on March 31, 2025, with access through its website and a new iOS app. The initial plans were credit-based, charging for the compute consumed by agent tasks.[10][18]
| Plan (launch, March 2025) | Price | Monthly credits | Concurrent tasks |
|---|---|---|---|
| Starter | US$39 / month | 3,900 | 2 |
| Pro | US$199 / month | 19,900 | 5 |
| Team | US$39 / seat / month | Per-seat allotment | 5-seat minimum |
The company later revised its pricing repeatedly as the product matured, adding a free tier with daily refreshing credits and restructuring its paid plans (for example, a lineup of lower-priced Pro tiers alongside a higher-priced power-user tier and team plans with administrative and single-sign-on features). Exact tiers and credit allotments have changed over time.[18][19] Some advanced capabilities were tied to plan level: Wide Research launched first on the top Pro tier before extending to lower-priced plans, and the Manus-1.5-Lite agent let subscribers run tasks at a 50% credit discount.[23][27]
How fast is Manus growing?
Manus scaled unusually fast for an agent product. In a December 17, 2025 update, the company said it had reached US$100 million in annual recurring revenue about eight months after launch, with a total revenue run rate above US$125 million once usage-based and other revenue were included, and month-over-month growth of more than 20%. It called Manus "the fastest startup to go from $0 to $100M in the world."[25]
The company also reported cumulative usage of 147 trillion tokens processed and 80 million virtual computers created, run by a team of about 105 people across Singapore, Tokyo, and San Francisco.[25] Reporting tied to the 2026 ownership dispute later described Manus's annual revenue as approaching US$500 million.[28]
Why did China block Meta's acquisition of Manus?
By late 2025, Butterfly Effect reported rapid revenue growth, with an annualized run rate that reporting placed around US$90 million in August 2025 rising to roughly US$125 million by December, and a headcount of about 100 to 105 employees across Singapore, Tokyo, and San Francisco. The company was reported to be raising a new round at a valuation near US$2 billion when, on December 30, 2025, Meta announced it had agreed to acquire Butterfly Effect. The price was not officially disclosed; outlets including reporting cited by CNBC and others put the deal at around US$2 billion (some accounts as high as roughly US$2.5 billion to US$3 billion), which would rank among Meta's largest acquisitions. Meta said it intended to keep operating and selling Manus and to integrate its technology into products such as Meta AI, while Manus said it would continue offering its own subscriptions and remain based in Singapore.[7][8][20]
The transaction quickly ran into Chinese regulatory resistance. United States Treasury scrutiny of the earlier Benchmark investment under the Outbound Investment Security Program (a regime requiring notification of certain US investments in AI and other sensitive technologies in "countries of concern") had already shadowed the company through 2025; Benchmark's outside counsel had argued the investment fell outside the rules because Manus did not train its own foundation models and Butterfly Effect was incorporated outside China. After the Meta deal was announced, Chinese authorities opened their own review: China's Ministry of Commerce began an assessment in January 2026, and in March 2026 co-founders Xiao Hong and Ji Yichao were summoned to Beijing and placed under exit bans that barred them from leaving the mainland, even as roughly 100 Manus employees had already relocated to Meta's Singapore offices.[21][22][26]
On April 27, 2026, China's National Development and Reform Commission blocked the deal under the country's foreign-investment security-review mechanism, sometimes described as China's counterpart to the US CFIUS process. Its statement said the NDRC "has made a decision to prohibit foreign investment in the Manus project in accordance with laws and regulations, and has required the parties involved to withdraw the acquisition transaction."[26][29] It was the first publicly confirmed use of China's Foreign Investment Security Review Measures to block and unwind a cross-border AI deal. A Meta spokesperson said the "transaction complied fully with applicable law" and that the company anticipated "an appropriate resolution to the inquiry."[26]
Meta formally unwound the acquisition in June 2026, completing an operational separation from Manus and halting data sharing between the two.[30] Reporting by The Information said Manus's early backers, including HongShan (HSG), ZhenFund, and Tencent, were preparing to buy the company back from Meta at roughly the US$2 billion price Meta had paid, with the founders seeking to raise about US$1 billion, while Benchmark was not expected to participate; Manus was said to be weighing a restructuring into a China-incorporated joint venture that could support a future listing in Hong Kong.[28] The episode was widely covered as a test of the limits of relocating a Chinese-founded company abroad, with one account observing that "the jurisdiction of incorporation didn't matter" because "what mattered was where the founders came from and where the technology originated."[12][30]
Related
References
- "Manus (AI agent)." Wikipedia, accessed June 4, 2026. https://en.wikipedia.org/wiki/Manus_(AI_agent) ↩
- "Manus: China's Latest AI Sensation." ChinaTalk, March 2025. https://www.chinatalk.media/p/manus-chinas-latest-ai-sensation ↩
- Kyle Wiggers, "Manus probably isn't China's second 'DeepSeek moment.'" TechCrunch, March 9, 2025. https://techcrunch.com/2025/03/09/manus-probably-isnt-chinas-second-deepseek-moment/ ↩
- Caiwei Chen, "Everyone in AI is talking about Manus. We put it to the test." MIT Technology Review, March 11, 2025. https://www.technologyreview.com/2025/03/11/1113133/manus-ai-review/ ↩
- Kyle Wiggers, "Chinese AI startup Manus reportedly gets funding from Benchmark at $500M valuation." TechCrunch, April 25, 2025. https://techcrunch.com/2025/04/25/chinese-ai-startup-manus-reportedly-gets-funding-from-benchmark-at-500m-valuation/ ↩
- "Manus revenue, funding & news." Sacra, 2025-2026. https://sacra.com/c/manus/ ↩
- "Meta acquires intelligent agent firm Manus, capping year of aggressive AI moves." CNBC, December 30, 2025. https://www.cnbc.com/2025/12/30/meta-acquires-singapore-ai-agent-firm-manus-china-butterfly-effect-monicai.html ↩
- "Meta Acquires Chinese AI Startup Manus for Billions in Third-Largest Deal." TechNode, December 30, 2025. https://technode.com/2025/12/30/meta-acquires-chinese-ai-startup-manus-for-billions-in-third-largest-deal/ ↩
- "How Yichao 'Peak' Ji became a global AI app hitmaker." MIT Technology Review, September 8, 2025. https://www.technologyreview.com/2025/09/08/1122642/ji-peak-yichao-innovator-manus-app-ai/ ↩
- "Manus AI Launches Paid Plans After Viral Demand, Offering $39 and $199 Tiers." WinBuzzer, March 31, 2025. https://winbuzzer.com/2025/03/31/manus-ai-launches-paid-plans-after-viral-demand-offering-39-and-199-tiers-xcxwbn/ ↩
- "People are listing invite codes to Chinese AI agent Manus for $1,000 and more on resale sites." Yahoo / Business Insider, March 2025. https://www.yahoo.com/tech/people-listing-invite-codes-chinese-185650467.html ↩
- "Manacled Manus: the limits of 'Singapore washing' for China AI." Asia Times, April 2026. https://asiatimes.com/2026/04/manacled-manus-the-limits-of-singapore-washing-for-china-ai/ ↩
- "China's Viral AI Agent Startup Manus Relocates to Singapore, Lays Off Domestic Staff Amid U.S. Tech Restrictions." TMTPost, July 2025. https://en.tmtpost.com/post/7620035 ↩
- "Manus AI lays off China staff, scrubs social media, shelves mainland service." South China Morning Post, July 2025. https://www.scmp.com/tech/tech-trends/article/3318310/manus-ai-lays-china-staff-scrubs-social-media-shelves-mainland-service ↩
- "Manus partners with Alibaba's Qwen to expand AI capabilities." TechNode, March 12, 2025. https://technode.com/2025/03/12/manus-partners-with-alibabas-qwen-to-expand-ai-capabilities/ ↩
- "GAIA: a benchmark for General AI Assistants." arXiv / Hugging Face, 2023. https://arxiv.org/abs/2311.12983 ↩
- "China's Manus AI Partners With Alibaba's Qwen Team in Expansion Bid." U.S. News / Reuters, March 11, 2025. https://www.usnews.com/news/technology/articles/2025-03-11/chinas-manus-ai-announces-partnership-with-alibabas-qwen-team ↩
- "What is the current membership pricing for Manus?" Manus Help Center, 2026. https://help.manus.im/en/articles/11711111-what-is-the-current-membership-pricing-for-manus ↩
- "Manus Plans & Pricing." Manus, accessed June 4, 2026. https://manus.im/pricing ↩
- "Review of Meta-Manus deal underlines China's tightening grip on AI exports." South China Morning Post, 2026. https://www.scmp.com/tech/big-tech/article/3339335/review-meta-manus-deal-underlines-chinas-tightening-grip-ai-exports ↩
- "The Manus Decision: China's First AI Security Review Block and Implications for Cross-Border AI Investment." Morgan Lewis, May 2026. https://www.morganlewis.com/pubs/2026/05/the-manus-decision-chinas-first-ai-security-review-block-and-implications-for-cross-border-ai-investment ↩
- "Why Did Beijing Kill a $2 Billion AI Deal?" Foreign Policy, May 6, 2026. https://foreignpolicy.com/2026/05/06/china-ai-singapore-manus-meta-us-competition/ ↩
- "Introducing Manus 1.5." Manus (official blog), October 16, 2025. https://manus.im/blog/manus-1.5-release ↩
- "Introducing Manus 1.6: Max Performance, Mobile Dev, and Design View." Manus (official blog), December 15, 2025. https://manus.im/blog/manus-max-release ↩
- "Manus Update: $100M ARR, $125M revenue run-rate." Manus (official blog), December 17, 2025. https://manus.im/blog/manus-100m-arr ↩
- "China vetoes Meta's $2B Manus deal after months-long probe." TechCrunch, April 27, 2026. https://techcrunch.com/2026/04/27/china-vetoes-metas-2b-manus-deal-after-months-long-probe/ ↩
- "Product updates." Manus (official), 2025-2026. https://manus.im/updates ↩
- "Manus original investors plan to buy back AI firm from Meta for $2 billion, The Information reports." U.S. News / Reuters, June 18, 2026. https://money.usnews.com/investing/news/articles/2026-06-18/manus-original-investors-plan-to-buy-back-ai-firm-from-meta-for-2-billion-the-information-reports ↩
- "China blocks Meta from acquiring AI startup Manus." NPR, April 27, 2026. https://www.npr.org/2026/04/27/g-s1-118892/china-blocks-meta-from-acquiring-ai-startup-manus ↩
- "Meta completes operational split from Manus, halts data sharing." Crypto Briefing, June 2026. https://cryptobriefing.com/meta-manus-operational-split-data-sharing/ ↩
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
5 revisions by 1 contributors · full history