We tested the “Chinese Claude Cowork”: the slap in the face is real, as is the price

Manus My Computer promises more autonomy and speed than Claude Cowork. We tested it on a concrete machine learning use case. Verdict: the agent impresses, but so does the prisoner.

It’s the AI ​​agent that caused SaaS stocks to lose billions of dollars in valuation. Launched in January by Anthropic, Claude Cowork tackled tasks that are difficult to automate for knowledge workers. With a specialization in the data aspect (thanks to a partnership with Microsoft on the Office suite), Claude Cowork quickly established itself as the reference agent for knowledge workers. Less than two months later, Manus, Anthropic’s most serious Chinese competitor on the agent side, unveiled Manus My Computer, a Claude Cowork clone with a positioning focused on autonomy and speed. However, is it as good as its competitor from San Francisco? Answer.

Manus, Claude: a different functioning

To use Manus My Computer, the Chinese publisher specializing in agentics has developed a Desktop application based on the principle of Claude Desktop. My Computer mode works more or less like Claude Cowork: the AI ​​agent takes control of your machine to carry out complex tasks. The Manus agent, confined like Claude Cowork in a folder, uses bash commands to analyze and modify files locally and can control your local applications with its computer use mode. Manus requests approval before performing an action, unless you grant it permission by default.

The application is available on macOS and Windows. Similar to Dispatch, Manus My Computer can be controlled from your phone with the Manus app or from Slack, Telegram or Line. Manus even goes beyond Claude’s capabilities and offers a turnkey API to create and launch tasks remotely. Manus highlights classic use cases (sorting photographs in a file, invoice management, etc.) but also more advanced cases. The Chinese publisher claims, for example, that it is possible to create applications from scratch with this mode. An employee of the start-up would have completely created a real-time translation and subtitles app in Swift in just 20 minutes. Manus markets its solution as an alternative to OpenClaw (without naming it by name) geared towards real use in businesses. The company therefore recommends installing My Computer on a dedicated Mac mini that is turned on 24 hours a day.

Manus Desktop is based on the family of agents developed by Manus: Manus 1.6. Three versions are offered:

• 1.6 Light for simple tasks
• 1.6 classic for tasks of medium difficulty
• 1.6 Max for complex use cases

Unlike Anthropic, Manus only controls his agent’s scaffold. If the start-up does not communicate the models used by its agents, the subcontractors declared on its Trust Center strongly raise our suspicions: Anthropic and Google Cloud in particular. We can therefore reasonably think that Claude is the driving model, supplemented by the Gemini suite on other tasks depending on the complexity. In any case, as all partners are based in the United States, your personal data should not be sent to China.

A credit price

Unlike Claude, billed at 20 dollars per month with Claude Pro or 200 dollars for the Max offer, Manus works on a credit system. Each action executed by the agent consumes a variable number of credits depending on its complexity: a simple task of sorting files will cost a few dozen credits, while a complex workflow (creation of an application, in-depth multi-source search) can swallow up between 500 and 900.

Three paid plans are offered: Standard at $20 per month (4,000 monthly credits), Customizable at $40 per month (8,000 credits) and Extended at $200 per month (40,000 credits). All plans, including free, get 300 daily refresh credits, access to all three versions of Manus Agent 1.6, and 20 concurrent tasks. A Team offer is also available with a pool of credits shared between members and administration functions (SSO, access controls). Annual billing offers a discount of around 17%. Manus’ pricing positioning is aggressive on paper (more common, in theory, than Claude Cowork), but the credit model introduces an unpredictability absent at Anthropic.

The JDN test

To concretely test the capabilities of the Manus agent, we gave it a machine learning challenge applied on a daily basis: building, from scratch, a model capable of automatically classifying files in the Downloads folder. The agent had to scan the folder, extract the metadata from each file (extension, size, timestamp, naming pattern), build a usable dataset, train a classifier to predict the destination category of each file (working documents, invoices, images, ephemeral files to delete, etc.), then apply the model to concretely reorganize the folder. All on a regular basis.

We thus give the following prompt to Manus Desktop with the most advanced agent (Manus 1.6 Max):

“Analyze my Downloads folder and build a machine learning model capable of automatically classifying my files into relevant categories. Start by scanning the entire folder to extract the metadata of each file (name, extension, size, creation date, last modification date, depth in the tree), then create a structured dataset from this data. Do feature engineering on the file names to deduce useful signals (presence of dates, hashes, keywords like invoice, contract, screenshot, IMG…). Trains a classification model (random forest or gradient boosting) to predict the destination category of each file (working documents, invoices and administrative documents, photos and images, installers and archives, temporary files to delete), evaluates its performance, then applies it to concretely reorganize my folder by creating the subfolders and moving the files. And sets up a sorting routine based on the model once a week.

Once the prompt has been sent and permanent permission to access the Downloads folder has been given, Manus launches. The agent begins by creating a plan. It then executes all the commands necessary to accomplish our use case in a virtualized environment (an Ubuntu VM). The agent consults the files, constructs a dataset and then trains a classification model. He finally ends up applying the model, produces our report and sets up the requested weekly sorting routine. First observation: Manus is much, much faster than Claude Cowork. The task only takes about ten minutes.

The final result meets our expectations: the product report (available here) is clean, structured, with model metrics and details of files moved by category. The classifier correctly identified the vast majority of our files and the tree structure created in the Downloads folder can be used immediately. On the credit side, the task consumed 1,078 credits, or more than a quarter of the Standard monthly package (4,000 credits) for a single operation. With just a few complex tasks per week, the credit budget can dwindle very quickly.

An excellent agent, despite unsuitable pricing

At the end of our test, the observation is clear: Manus My Computer is, on paper, an excellent alternative to Claude Cowork. The Chinese agent does the work, and he does it quickly, very quickly in fact. The real downside is the consumption of credits. With intensive use, the bill can very quickly exceed that of a Claude Pro subscription at $20, or even rival the Max offer at $200, without the predictability offered by a fixed plan.

For companies that really want to automate their processes, Manus still has major advantages: its turnkey API, its architecture designed for deployment (dedicated Mac mini, remote control) and its almost total autonomy. We now hope that Manus will review the pricing… Unless this pricing (increasingly common -> Devin link) simply reflects the real cost of agentic AI in 2026, which no one really wants to admit yet.