🐙 What's the deal with Manus?
Six things you need to know to understand the hype
The online discourse around Manus AI typically falls into three camps:
The enthusiasts – “Just tested Manus AI and my jaw is on the floor.”
The impressed professionals – “Manus is the most impressive AI tool I’ve ever tried.”
And the skeptics – “Manus AI is a fraud, it’s just a Claude Sonnet MCP wrapper”.
There’s a wide range of opinions about Manus. We’ve tested it out – here are six things you need to know about it.
1. What is Manus?
Manus is a general-purpose AI agent. You give it a task—say, analyzing Tesla’s stock—and it crafts a plan to handle that step by step. It can browse the web, write code and tap into various tools to get the job done.
In my experience, Manus is seriously impressive and far outperforms OpenAI Deep research on two of the three tasks I gave it, including producing a detailed itinerary for a holiday and evaluating changes in the media landscape. Deep research qualitatively felt better for a task evaluating the various chip architectures.
How to get it: Currently, you need an invitation code. You can sign up for the waitlist on Manus’s official website.
Cost and availability: It’s free during beta (with usage caps). Once it launches, it may not be significantly cheaper than other providers.1
Note: Video replay is sped up. You can see the dashboard here.
2. Why would you use it?
An early experiment with agents: Using Manus will show you how AI agents could automate research or planning tasks – and how you can already start relying on them to automate parts of your work. Manus isn’t ready for full-scale business use. Instead, test it out to see where AI-powered agents are headed.
Similar AI tools exist, but Manus offers better execution: Manus is in the same category as Claude Computer Use or OpenAI Operator and to a lesser degree as Deep research. Similar open-source AI agents (like Owl) exist, but they do not yet match Manus’s polish.2 According to benchmarks, Manus slightly outperforms OpenAI’s Deep research on the GAIA agent test. In my testing, Manus handles tasks a bit better than those other products—though it can still make mistakes or miss obvious details.
Use cases: Try it for research summarization, comparative product analysis, generating viral snippets from long videos, or even generating fun retro-style (SNES-era) games.
Data and privacy: The company is effectively Chinese-owned, although it maintains a legal entity in Singapore. Be cautious when sharing sensitive data.
3. It’s a “wrapper”
Some critics call Manus a “wrapper” because it doesn’t rely on a proprietary model.
Instead, it layers software around existing AI models—namely “Claude Sonnet 3.5” and fine-tuned “Qwen.”
The critique: Skeptics argue the true power lies in these foundational models, not in Manus’s interface or workflow design.
Pricing implications: Since Manus depends on third-party models, it may struggle to undercut pricing from services like OpenAI. OpenAI’s Deep research costs $20/month for a limited quota, while Operator is $200/month for pro.
4. That doesn’t mean it isn’t good
Despite using older models, Manus has delivered impressive results:
29 integrated tools: Manus has access to tools like “browser_navigate,” “shell_exec,” and “file_read” that help AI agents interact with websites, manage system processes and manipulate files effectively.
Superb integration of other AI startups: Much of Manus’s edge comes from building on top of specialized platforms. It uses Browser Use,3 a startup that helps AI agents interact more effectively with websites, and runs its sandbox environment through E2B, an open-source runtime for secure cloud execution of AI-generated code.
Engineering effort: Co-founder Yichao Ji cites inspiration from “CodeAct,” a system enabling AI to generate Python code as a universal problem-solving approach. The Manus team’s “secret sauce” likely involves how they fine-tune Qwen models, optimize context length and orchestrate planning.
Performance: It already performs at a high level without tapping the latest models (e.g., Claude Sonnet 3.7 or QwQ-32B).
5. More is on the way
Open-source enthusiasts have tried to replicate Manus—some even appear to have successfully obtained Manus’ code from Manus itself. We attempted verification but discovered the code was encrypted with PyArmor on March 7th (two days before the March 9th tweet). The initial release occurred on March 6th, which means there may have been a window of opportunity for this to happen.
Alignment vs. capability: In his conversation with HuggingFace’s Clement Delangue, Manus co-founder Yinchao ‘Peak’ Ji suggested that the challenge is more about “alignment”—training models to break tasks down properly—than raw power. Given this, we’ll likely see numerous AI agents emerge soon, as startups can easily implement different alignment approaches by building on existing foundation models.
Chinese partnership: Manus recently announced a strategic cooperation between its Chinese Edition and Qwen (Alibaba). The two teams want to bring all Manus features to domestic (Chinese) models and computing platforms.
6. Is there a moat?
Probably not yet: Manus’ advantages – fine-tuning of Qwen models, optimizing context length and orchestrating planning – might be non-trivial to replicate, but not impossible. We are also so early into agents that better methods are likely to emerge.
Fierce competition: Even OpenAI is working to defend its lead, as open-source and Chinese models rapidly improve. OpenAI agents are no longer everyone’s default choice.
Sign of what’s next: Manus demonstrates how quickly a new entrant can impress by combining proven large language models with good engineering.
Bottom line: Manus is not the final word in AI agents, but it’s a strong indicator of where we’re headed: a future with more autonomous AI helpers—and fast. Keep an eye on it, as it’s the sign of a coming wave on the horizon.
Under the hood, Manus is using Claude 3.5 Sonnet ($6 per million tokens) and fine-tuned Qwen models (likely $0.8 to 2.8 per million tokens). Meanwhile o3-mini is running at $1.93 per million tokens.
Deep research is more focused on performing research tasks; Manus can do more general tasks, like generating a dashboard.
This startup is worth checking out if you're looking to build web agents.
Thank you for your analysis as always. I appreciate you as a trusted guide in all of this. Thanks.
Great overview. May these tools be used for more than frenzied shopping. The biosphere can’t handle much more business as usual on autopilot.