In an unassuming office in Beijing, a developer types a simple prompt: "Create an interactive dashboard for Tesla stock with a forecasting model." Then he sits back and watches as the AI takes over. Within minutes, a browser opens, data is retrieved, code is written, and suddenly a fully functional dashboard appears, complete with stock prices, forecasting models, and interactive elements—all without further human intervention.
Welcome to the world of Manus AI, the latest wunderkind from China's booming AI scene, which has been stirring up excitement in the tech world since its launch on March 6, 2025. While OpenAI, Anthropic, and Google are still tweaking their chatbot interfaces, Chinese startup Monica (also known as Butterfly Effect) claims to have created the first "fully autonomous AI agent" with Manus—a claim as bold as the coveted invite codes, which now sell for thousands of dollars.
The second “deepseek moment”?
In early 2025, DeepSeek rocked the Western AI world. The Chinese startup had developed a language model that could compete with the best Western technologies—and at a fraction of the cost. The announcement triggered a $600 billion drop in value for Nvidia and demonstrated that China was capable of developing leading AI models despite US export restrictions on advanced chips.
Just two months later, the pattern appears to be repeating itself. Manus AI is attracting attention with its claim of complete autonomy, contributing to the narrative that China is rapidly catching up in AI development.
"DeepSeek was about replicating capabilities that American companies had already achieved. Manus is actually pushing the boundaries. The most advanced AI system now comes from a Chinese startup, period," explains Dean Ball, an AI policy analyst who is closely following developments.
From chatbots to real agents
Yichao "Peak" Ji, co-founder and chief scientist of Monica, emphasizes in his introductory video: "This isn't just another chatbot or workflow. It's a truly autonomous agent that connects ideas with execution."
This claim initially sounds like standard Silicon Valley marketing, but the demos the company has released are impressive. Manus can not only answer questions but also independently solve complex tasks—from screening 1,200 resumes in 18 minutes to creating dynamic financial analysis dashboards with machine learning-based forecasts.
The user interface is remarkably simple: You describe what you want to achieve in natural language, and Manus processes these instructions into complete results. In one demonstration, the agent analyzed Tesla stock and automatically created an interactive dashboard, which was published to a public URL.
The technology behind the hype
While Manus AI is marketed as a revolutionary product, its technical architecture is actually a clever patchwork of existing components. The system uses a multi-agent architecture that breaks complex tasks down into smaller sub-steps and solves them sequentially.
According to research and statements from the company, Manus combines the following technologies:
- Claude 3.5 Sonnet of Anthropic as a central language model
- Fine-tuned Qwen models from Alibaba for specialized tasks
- Browser Use – Open source software for web interaction
- In-house developed multi-agent architecture with dedicated “Planner”, “Knowledge” and “Executor” components
Peak Ji later confirmed in a post on X: "We are using Claude and various Qwen fine-tunings. When we started developing Manus, we only had access to Claude 3.5 Sonnet v1, which required the use of numerous auxiliary models. Claude 3.7 looks very promising; we are currently testing it and will provide updates!"
The system’s workflow follows an iterative agent loop:
- Analyze events: Processes user requests and the current task status
- Select tools: Select the appropriate tool or API for the next step
- Execute commands: Runs shell scripts, web automation, or data processing in a Linux sandbox
- Iterate: Refines actions based on new data
- Submit results: Sends structured output to the user
- Standby mode: Waiting for further user input
Each Manus session runs in an isolated sandbox environment with access to 29 integrated tools, from code editors to data visualization suites.
Performance and benchmark claims
Monica claims that Manus passed the GAIA benchmark—a benchmark for general AI assistants—with an impressive 86.5% compared to 67.9% for OpenAI's Deep Research feature. Internal tests also show:
- 92% Success rate for multi-stage research assignments
- 87% Accuracy of financial forecasts over 6 months
- 5x faster task completion times compared to Claude 3.5 Sonnet in standalone mode
However, independent verification of these performance data is still lacking, and critics such as AI ethicist Dr. Lena Schröder of the Technical University of Munich warn: "Benchmarks alone do not constitute AGI. The real question is: Can Manus truly solve novel problems or just extrapolate trained patterns?"
The two sides of the hype
Reactions to Manus in the tech community are mixed, ranging from effusive praise to deep skepticism.
The enthusiasts
Some prominent voices in the tech world are enthusiastic about Manus' capabilities. Victor Mustar, product lead at Hugging Face, described it as "the most impressive AI tool I've ever seen" and suggested that its capabilities could redefine programming: "This could eliminate traditional coding... it's more about conceptualizing ideas."
A beta tester from the financial industry reports: “Manus has reduced our due diligence processes from 40 to 6 hours” – albeit with the caveat of a “121% error rate in complex contract analyses.”
The skeptics
On the other side are the critics who see Manus as an overhyped product. One Reddit user disparagingly compares it to failed hardware projects: "It looks like it's about to be renamed Manure AI, essentially a second version of Rabbit.ai, inspired by Humane Pin."
Another user critically remarks: "It's Claude who uses tools. These tools are straightforward—quite simple, in fact. They've achieved something quite simple with a model that doesn't belong to them. Without real innovation, the landscape remains unchanged."
MIT Technology Review gained access to Manus and found that while the system is promising, it suffers from "frequent crashes and system instability" and "has difficulty processing large amounts of text."
The invitation code economy
A central aspect of the Manus phenomenon is its artificial scarcity. The system is invitation-only, which has led to a veritable gold rush. The official Discord server has grown to over 170,000 members, but fewer than 11 percent of users on the waitlist have received an invitation code.
This strategy of artificial scarcity has created both a viral marketing effect and a shadow economy. Invitation codes have reportedly been traded for thousands of dollars on China's resale platform Xianyu and on eBay.
Pierre-Carl Langlais, co-founder of the AI startup Pleias and an early user of Manus, criticizes this "misleading communication" and "starvation marketing" strategies, which he argues create artificial hype by limiting access to a select group of influencers. "What the AI industry really needs now are improved standards of transparency and openness at all levels: model, data, and business," Langlais emphasizes.
Implications for global AI competition
Manus' rapid rise has raised alarm bells in Europe and the US. Greg Nieuwenhuys, senior partner at the Amsterdam-based consulting firm Generative AI Strategy, warns: "Manus represents a major step forward in AI autonomy and marks another breakthrough from China."
"Currently, Europe does not have an equivalent AI agent project with the same level of autonomy. This raises concerns about whether European AI companies can remain competitive in the long term," adds Nieuwenhuys. "Without strong investment and government-supported initiatives, Europe risks being left behind in the race for AI leadership."
While DeepSeek focused primarily on improving the capabilities of large language models, Manus bridges the gap between thought and action. It doesn't just process information—it acts on it, automating tasks from start to finish.
This shift signals that China is moving from AI assistants with large language models to AI agents capable of autonomous execution—something even OpenAI has not yet fully mastered.
Roadmap and challenges
While Monica has ambitious plans for Manus, the company faces significant challenges:
Short-term hurdles
- Scaling: The current architecture only supports 10,000 parallel sessions
- Cost: $3.80 per task in high performance mode vs. $0.80 on OpenAI
- regulation: China's new AI audit requirement from June 2025
Long-term vision
Ji outlines the goals in an X-Space interview:
- 2025: 100,000 active corporate customers
- 2026: Integration of physical robots via ROS 2
- 2027: “Manus OS” as AI operating system
The company also plans to open source parts of its system, including the sandbox controller (Q2 2025), custom Qwen finetunes (Q3 2025), and the tool integration API (Q4 2025).
Conclusion: Revolution or hype?
While the servers in Beijing are running hot, the most burning question remains: Is Manus really the long-awaited AI quantum leap or just a cleverly choreographed machine learning ballet?
The race to create the first true AGI has undoubtedly opened a new chapter with Manus. Its ability to perform complex tasks autonomously is impressive, but the exaggerated marketing claims and technical challenges dampen the enthusiasm.
What distinguishes Manus from other AI systems isn't necessarily a revolutionary new architecture, but rather the clever integration of existing technologies into a coherent agent system. It leverages the strengths of Claude and Qwen, combines them with a multi-agent architecture, and adds an execution layer that enables autonomous actions.
As with all AI startups, there's often a gap between ambition and reality. The truth about Manus lies—as so often—somewhere between the ecstatic claims of its proponents and the skeptical objections of its critics.