Last month, a new Chinese AI began making headlines. Called Manus, the AI presents itself as a "general AI agent" capable of connecting "the thoughts and actions" of its users. On its website, the Chinese startup Monica, behind Manus, assures that its AI agent "doesn't just think, it delivers results." Manus excels "at various work and life tasks, accomplishing everything while you rest." On paper, Manus is selling a dream. Mirroring DeepSeek, Manus has therefore been much talked about on social media.
Manus is part of the trend of AI agents, that is, artificial intelligences designed to assist users by demonstrating autonomy and initiative. The robot must be able to complete a project from A to Z, performing all the necessary tasks without having to rely on humans. To achieve this, it uses a series of AI models, including Anthropic's Claude 3.7 and Alibaba's Qwen models. In theory, Manus represents the future of artificial intelligence. After all, we've all dreamed of having an assistant capable of taking care of all the menial and tedious tasks for us.
Intrigued by the promise of Manus, we quickly joined the waiting list set up by the Chinese startup. After a few weeks of waiting, we received access to the free version of Manus. We rushed to test the AI agent to see if it lives up to its promises. Our test was conducted using the Manus macOS app. From time to time, we also used the iPhone version to take a look at the results of the generative AI.
A very limited free version
First observation: the free version of Manus is considerably limited. Indeed, the Chinese company only offers you 1,000 credits to test its intelligent agent. As we noticed, you quickly run out of credits. In fact, you only have to make three slightly advanced requests for Manus to inform you that you have no more credits. The startup specifies that "credits are our standard unit of measurement for using Manus - the more complex or long a task is, the more credits it requires". Good news, Manus offers "a full refund of credits consumed for tasks that fail due to technical issues on our end.".
For the purposes of testing, we decided to subscribe to a paid subscription. Manus offers two different plans, at very different prices. First, there's the Starter subscription, which costs 45 euros per month. This standard offer includes 3,900 usage credits, which represents between 15 to 40 tasks per month, depending on their complexity. It all depends on what you ask of Manus.
The subscription also promises "improved stability thanks to dedicated resources, extended context length, and priority access during peak hours." You can also run two tasks simultaneously, which quickly becomes essential. Indeed, some of the requests sometimes take a long time. We're often tempted to ask AI simpler questions while another, more complex problem is being addressed.
Manus also highlights a Pro subscription… priced at €229 per month. This package offers 19,900 credits per month, or between 50 and 200 tasks. In addition to the benefits already offered by the Starter offer, the subscription allows you to run five tasks simultaneously.
Mixed results
To begin, we asked Manus to develop a website from A to Z about the data leaks that occurred in France. To create this site, the AI was forced to draw from a multitude of online sources and compile them into a coherent whole. We were amazed by Manus's ability to understand a problem, structure its work, retrieve data, and code a website. The developed site was fully functional and visually correct. The operation took no more than ten minutes. We invite you to take a look at the finished product by clicking on this link.
For all requests, Manus clearly displays the progress of all tasks completed to complete the project. To create the site, Manus began by researching information before creating the site structure. He then wrote all the site pages, followed by the HTML and CSS code. The finished product was communicated as an HTML file. The AI keeps you informed of its progress in real time. While he's working, you can close the window and go about your business.
We were a little disappointed when we asked the AI agent to design a document based on a series of files we provided. It seemed to us that Manus doesn't have a deep understanding of the French language. The way Manus formulated the requested document was very mechanical and not very fluid. We're a long way from the prose of ChatGPT, Grok, or even Claude. Manus had difficulty adapting his writing style to our requests. We told the AI that we needed to talk to children to test how it would correct the situation. The results weren't very conclusive. Writing isn't really his thing. We were actually surprised by some of the wording, such as "I have attached a detailed document with a complete budget breakdown."
On the other hand, Manus was very impressive in tasks requiring digging into external resources. We asked the AI to draw up an itinerary for a trip, including activity recommendations, a complete program, and gastronomic advice. In a short time, Manus created a PDF summary of our trip, including a map of nearby restaurants, the schedule for the chosen hotel's spa, the location of electric car charging stations, and the prices of all nearby activities.
In a few minutes, we had a real map offering an overview of everything we could do during our trip at our fingertips. No need to go to a search engine to retrieve information on our own from multiple sites. It's a serious time saver. We obtained similar results by querying ChatGPT. However, OpenAI's AI was less precise in its responses. Manus is much more methodical and rigorous. Unlike ChatGPT, Manus is used to asking for clarification before embarking on a project. For example, the AI asked us these questions before starting work on our travel program:
We also used Manus to obtain an accurate estimate of the budget needed for a trip, without having to do any calculations. Here again, Manus stood out for its accuracy and a simple and readable presentation of the data. The chatbot systematically offers to convert its production into a website, easy to open in a browser. The speed with which Manus produces a website makes it easy to navigate through the data. It's also very practical if you need to share the information with third parties, for example in preparation for your trip. You are free to keep this site private, that is, to protect it with a password, or to make it publicly accessible, to anyone who has the URL designed by Manue. This time, the website was visually more sophisticated, without frills or bad taste. Manus explains that he integrated images of our upcoming trip to "make the site more attractive". We invite you to take a look at the site created with Manus to get an idea of what it's capable of.
Manus excels especially when it has to juggle information contained in resources provided by us. For example, we asked Manus to synthesize PDFs, Excel files, and emails received from an accounting department, and the results were there. It's a real asset in administrative tasks that require patience, organization, and discipline.
We really appreciated the fact that Manus explicitly displays what it's doing. This feature allows you to understand how the AI process works, and to intervene in real time to correct it in case of misunderstanding. It's by closely monitoring Manus that we obtain the most calibrated results from the first response.
A good data collector, but a poor writer
Manus seems gifted at retrieving data, online or from documents, and aggregating it in one form or another. On the other hand, the AI agent seems to lack creativity and finesse. In fact, we really feel like we're interacting with a robot. We are obviously aware that it is indeed a robot. This feeling is nevertheless absent, or less present, in our daily interactions with generative AIs, like ChatGPT and others. Generative AIs were programmed to communicate as naturally as possible. This is why Manus's cold and mechanical responses surprised us.
In this respect, Chinese AI is very different. We immediately feel that we are interacting with a computer program, which has its own limitations and its own functioning, rather than with a true assistant capable of adapting to human reasoning. It is possible that Manus will evolve and improve on this point in the future.
Another problem encountered during our experiments: the files sent by Manus are sometimes difficult to download. Once the AI has completed its task, it sends you a document that lists all its research and achievements. On macOS or iOS, we often had trouble retrieving the document. We had to close the application and restart it. It's a detail, but it was recurrent enough that we should mention it in this test. More generally, it often happened that a particularly long and complex operation crashed due to a bug. The system often appeared very unstable, whether on the macOS app, on the iPhone, or in the browser. We imagine that the Chinese startup's servers were a little saturated.
Is Manus really revolutionary?
Ultimately, we found Manus terribly promising. For certain tasks, the AI works like a true intelligent assistant. It stands out for its impressive autonomy and its ability to combine a mountain of tasks to achieve the expected result. On the most complex operations, Manus is more efficient than ChatGPT. More precise, more methodical, and more rigorous, it carried out tedious tasks without our help, and saved us a lot of time.
Nevertheless, the AI agent is far from perfect. Over the course of our experiments, we realized that Manus doesn't always fully understand the nuances of what we ask of him. To show his best, Manus needs very precise and well-structured information. Don't leave room for interpretation. This is a recipe for failure. Manus isn't always able to correctly interpret the nuances or intentions behind your requests. Speak to him like you would a somewhat obtuse intern. Be factual and precise. When properly calibrated, Manus becomes a solid assistant, whether for research, organizing daily life, or administrative management, but he absolutely needs your guidance.
0 Comments