Written by | Hao Xin

Edited by | Wu Xianzhi

Since yesterday, Manus has descended upon the AI community, dominating the entire internet.

Despite the tidal wave of popularity, Manus has found itself mired in polarized controversy. Some have cheered it on, hailing it as the "next DeepSeek moment" that "overnight revolutionized OpenAI and Anthropic." Others have criticized Manus for being all talk and no action, with its hunger marketing through "invitation codes" being off-putting. In actual evaluations, the performance of running Agent cases leaves much to be desired, with common issues including long processing times, lag, and repeated outputs.

The current situation is that free invitation codes have been inflated to astronomical prices. It is reported that the Manus team has found itself in an unexpected predicament, with the overwhelming popularity causing their servers to crash.

For unknown reasons, Manus' official account on the overseas social media platform "X" has been frozen.

Amidst the scattered and fragmented information, we attempt to first salvage some key facts about Manus.

Manus is officially positioned as a "General AI Agent," a multi-model, multi-agent product with the ability to "think independently," capable of breaking down and executing complex general tasks in multiple steps.

Manus advocates the philosophy of "Less structure, more intelligence," where the Agent's ability evolves based on improvements in the underlying model's capabilities and the increase in data, achieving natural evolution rather than enhancement through workflows.

In the demo presentation introducing the product positioning, Manus is not just another chatbot or workflow but a truly autonomous entity that bridges the gap between concept and execution. While other AIs are merely generating ideas, Manus is delivering results.

Manus candidly admits that its product is a "shell," with the underlying models being Claude and Qwen, and the cost for a single task consumption being approximately two dollars.

"Toy" for Agents

Manus' definition of Agent products has returned to a more pure level, allowing Agents to complete tasks independently without human intervention. The concept has been around since AutoGPT but has been hindered by the capabilities of the underlying large models, necessitating the use of external tools and prescribed workflows to ensure Agent execution effectiveness.

One of the most impressive aspects of the Manus demo is that it truly achieves "what you see is what you get," which is the most essential delivery logic for Agents. For example, previously, one had to analyze files step-by-step before creating a table, but now it can be done in one step. After identifying user needs, it executes the steps and directly delivers the final table.

In its officially released demo videos, Manus can analyze resumes, create tables, produce real estate research reports, filter out the best options based on budget, analyze stocks, and more.

As of now, the above functions can be achieved through any AI assistant application. However, in a sense, Manus can be considered the first complete Agent product. While OpenAI, Anthropic, and Zhipu have all released Agent-related functions, most exist as tool attributes.

Upon opening the Manus interface, it consists of four parts: browser, search, editor, and terminal. Manus' self-proclaimed "shell" status is well-deserved, as it is a product that encapsulates and combines various functions, similar to the AI search engine Perplexity.

There's a side note: it is said that the Manus team initially aimed to develop an AI browser. After stumbling upon the abandonment of Arc, they decided to terminate their AI browser development efforts. Traces of the browser can still be seen in its current usage.

Specifically, the first step is similar to ChatGPT, with a dialogue interface where users need to propose specific requests. The operating interface is then divided into two parts: the left side is the dialogue interface, and the right side is the terminal.

When a task is initiated, the left side begins to recognize intent, formulate execution steps, and start searching and invoking various required tools. The terminal on the right side acts as a virtual machine that can synchronously execute file processing, code generation, search browser operations, etc. The so-called Computer Use mainly involves simulating user clicks, browsing, and switching tools, which has limited significance.

In this way, Manus is roughly equivalent to "ChatGPT/Claude/DeepSeek Model + Agent Framework + Search Engine + Tools + Computer Use + RAG." The team's innovation lies in utilizing many engineering methods to combine the above logic seamlessly like building blocks.

Judging from the demos currently shown, Manus is well-suited for C-end users due to its low threshold and concise, straightforward functionality. If the effects are genuine, it can still provide a good user experience.

According to the "AutoHua.AI" public account, Manus mentioned the "three axes" to enhance the future user experience:

1. Configuring computers to give AI access to browsers and tools, such as cloud browsers. 2. Opening permissions to access private APIs and authoritative data sources, such as financial indicators. 3. Dynamic training, where users can adjust AI behavior in real-time through feedback, similar to training Agent interns, adapting to user needs after a few days of use and enhancing the user experience.

Changing the Supply and Demand Relationship

Manus' various marketing tactics are essentially attempts to preemptively bet on the Agent application race. Through Manus, we can see the shift in supply and demand relationships in the era of large models.

Agents ultimately aim for delivery, speaking with efficiency and results. This means AI becomes the service provider, and humans become the consumers. Demand determines supply, and supply meets demand. Once the logic of the market economy can form a closed loop in Agent applications, it can promote consumer behavior and ultimately realize value.

Previously, many people hoped that AI search could fulfill the above mission, but even the leading company Perplexity only thought of "a good way" to place advertisements. Now, Agents are moving towards productivity.

Manus' functionality easily evokes comparisons to OpenAI's Deep Research. It is also an Agent that can research, think, and complete reports like an experienced researcher. OpenAI officially claims that this feature reduces tasks that would originally take humans 8 hours to just 5 minutes, helping people save hours or even days at work.

According to the latest news, OpenAI plans to launch a tailored Agent for professionals to perform advanced tasks such as sales lead classification, software engineering, and doctoral-level research.

When tied to productivity, commercialization becomes natural. Based on current thinking, there are roughly two categories: subscription-based charging and task-oriented results-based charging.

OpenAI's Agent service adopts a subscription-based charging model. It is reported that in the future, Agents will serve as the primary engine for revenue growth.

The Agents tailored for professionals are divided into three tiers: the first tier is for high-income knowledge workers, charging $2,000 per month; the second tier is for software developers, charging $10,000 per month; and the third tier is for doctoral-level research, charging $20,000 per month.

Based on Manus' ideas, it is highly likely that they will explore the second option, a task-oriented charging model.

At the sharing session, Manus introduced the concept of "Agentic Hours per User (AHPU)," which measures the time efficiency of users delegating tasks to AI, with the goal of enhancing productivity through parallel tasks. Currently, it has achieved a cost of approximately two dollars per single task through KV cache optimization, inference latency compression, execution process streamlining, and other means.

This provides another approach where users with low usage frequency can be charged based on single tasks. As long as users can obtain value and satisfaction from AI, they can complete the payment behavior.

Who are the Ultimate Beneficiaries?

Of course, the Agent capabilities and charging models discussed above are still in a very idealized state.

As we all know, "shell" products do not have a moat. But who would have thought that it would collapse overnight. Yesterday, Manus "blew up the scene," and today, the open-source community is reproducing it by frame-by-frame analysis of videos. The "CAMEL AI" public account published a post today titled "Replicating Manus' General Intelligent Agent in 0 Days, Completely Open Source," breaking down Manus' core workflow into six steps.

One Manus has vanished, but countless Manuses have emerged.

The enhancement of Agent capabilities still heavily relies on the capabilities of large models. Manus' current approach is to piece things together, using one large model to compensate for the inadequacies of another. Although it promotes itself as "Agent native," given its team size and training capabilities, it cannot be ruled out that there is still a possibility of building workflows. In the current context, more workflow designs represent stronger controllability. Overall, most players in the industry are at this level.

On the other hand, the Agent released by OpenAI follows a completely model-training path. A prominent feature of Deep Research lies in the autonomous ability evolution brought about by end-to-end training. Based on the fine-tuned O3 version, the underlying training endows Deep Research with many analytical capabilities. In the long run, reinforcement learning adjustments on top of the model may be the key to building powerful Agents.

In the cases provided by Deep Research, it has already covered travel planning, stock analysis, supplier procurement, educational content creation, online store operation analysis, etc., basically covering Manus' functionalities. In the short term, if Manus wants to stand out, it will have to rely on its user experience and low-price strategy. In the long run, there is a risk of being overshadowed by OpenAI, as a "terminal" design poses no threat to it.

Cost may be the most critical issue. Manus claims that the cost for a single task completion is approximately two dollars, translating to hundreds of thousands to millions of token consumption. Considering the complexity of user demands, each single task execution may also involve supplementary task requirements, posing significant challenges to server resources and computing power. Reports indicate that an Agent task may consume 10-100 times more resources than a traditional chatbot.

However, this is not an issue for large companies. Yuanbao, by integrating DeepSeek, has come from behind, with its in-app inference functionality being smoother than the official DeepSeek. Large companies now almost universally have their own AI applications, self-developed large models, and access to the DeepSeek-R1 inference model. RAG and online search are readily available. For them, rebuilding or modifying an existing system to create a Manus-like product is hardly a problem.

If considering in-situ modifications, a tiered payment model might be considered. Adopting a free strategy for most ordinary users can attract new users and expand the user base. On this basis, high-quality paying users can be screened out, with tiered pricing similar to OpenAI, where the higher the service quality provided, the more tokens consumed, and the higher the fee charged.