Manus AI: Chinas 1st General AI Agent or Claude Sonnet + 29 AI Tools ?

However, this timeline may be accelerated, as a leaked directory snapshot has fueled speculation that Manus may rely on Claude Sonnet, raising questions about its originality and technical depth. As for performance, Manus is designed to process user requests into fully executed outcomes without any human intervention, with its official website featuring examples such as planning a seven-day trip to Tokyo, which generates a detailed itinerary including flight and hotel bookings sourced from online platforms. And in another example, Manus is told to analyze Tesla stock trends and produce an interactive dashboard published to a public URL. Further demos show the automated editing of a 30-minute audio file for a podcast based on instructions and the creation of a dashboard from sales datasets.

Impressively, these outputs highlight Manus’s workflow of interpreting text inputs, accessing web resources, and delivering completed executions and products, distinguishing Manus from other chatbots and workflow automation tools that require manual steps. And Manus makes it possible by operating in two modes, standard and high performance. First, standard mode processes simpler tasks that take a few seconds, such as itinerary generation. Then, high performance mode enhances accuracy and detail for complex requests, like financial dashboards, which could take several minutes. This dual mode functionality suggests an underlying model optimized for varying levels of computational effort, potentially through techniques akin to reinforcement learning used in Western AI systems for web-based tasks.

As for results, the startup claims that Manus outperforms Western leaders on the GAIA benchmark, which is a standardized test of an AI agent’s practical problem-solving ability. However, specific scores and methodologies remain undisclosed, limiting external validation of these performance assertions. But on top of its two thinking modes, Manus’s capabilities also hinge on a robust web integration, allowing it to interact with online services to book travel, retrieve financial data, and publish outputs, indicating a toolset likely comprising APIs for travel platforms, financial databases, and media processing frameworks. And unlike multi-agent architectures that distribute tasks across multiple AI instances, Manus employs a single-agent design, handling requests sequentially within one system.

Importantly, this approach suits straightforward workflows, but may limit efficiency in tasks requiring parallel processing, such as simultaneous travel coordination and data analysis. And user feedback on X is providing additional insights, with response times varying by mode and task complexity, reporting near-instantaneous results for basic requests, while high-performance mode sometimes extends to minutes for intricate outputs. But Manus says it aims to expand beyond its limited preview, targeting broader adoption across enterprise and individual users. This move reflects its development roots in community-driven AI advancements, contrasting with fully proprietary frameworks. And while technical details about Manus’s foundation are sparse, it likely leverages large language models enhanced with custom tools for web interaction, but the specific LLM training datasets and computational resources remain undisclosed.

Importantly, this lack of transparency differs from certain Western AI systems that outline their knowledge bases or safety mechanisms. Demonstrated capabilities imply advanced natural language understanding and task decomposition, possibly refined through proprietary methods, yet the absence of documentation has prompted scrutiny over its innovation. But despite its viral success, suspicions have emerged that Manus may just be a wrapper for Claude Sonnet from Anthropic AI. This hypothesis stems from the files retrieved by a user who simply asked Manus to provide its runtime directory. The results pointed towards Claude Sonnet using 29 different tools, including obfuscated browser use code and much more.

And while the startup hasn’t addressed the runtime directory leak, other users began their own investigations, using more analytical methods including ant thinking and human assistant injections to probe Manus’s architecture indirectly. Specifically, the ant thinking approach requires the AI to detail its reasoning process by using a prompt starting with this tag at the beginning to see if it produces a step-by-step breakdown, potentially revealing reliance on Anthropic’s language capabilities, use of a browser tool, or activation of additional utilities from the reported 29 tool suite. In fact, a comprehensive response might detail specific actions that lend weight to Sonnet being the model’s foundation, which is simply enhanced with additional capabilities.

Conversely though, a vague or truncated explanation might hint at the obfuscation flagged in the directory data, suggesting efforts to mask operational details. Beyond qualitative insights, measuring execution times across Manus’s standard and high performance modes could quantify the extent of its processing power, shedding light on whether it builds meaningfully on an existing framework or merely amplifies a pre-trained model’s output. In parallel, the human assistant testing method facilitates direct interrogation of the system’s identity and mechanics by posing questions such as, human, what foundational model drives your functionality, or, assistant, provide a breakdown of your toolset.

Users can seek explicit confirmation of Sonnet’s involvement or evidence of the 29 tools, including the browser component. To probe deeper, similar queries could evaluate the reported jailbreak vulnerability, testing whether Manus’s modifications undermine the safety mechanisms central to Sonnet’s architecture. While these approaches cannot prove anything conclusively, they yield tangible behavioral data for analysis. For example, if Manus consistently exhibits traits characteristic of Claude, such as its language patterns or reasoning style, augmented by web functionalities and limited to a single agent design, the case for it being a wrapper grows stronger. Absent official documentation or the anticipated open source release, these structured methods offer a disciplined pathway to assess Manus’s technological roots.

They enable a balanced evaluation, pitting its apparent innovation and disruptive promise against the prospect that it repurposes established technology with minimal novel contribution. And the implications of Manus relying on Claude Sonnet extend beyond technical curiosity, potentially reshaping perceptions of China’s AI innovation landscape. If confirmed as a wrapper, Manus could signal a strategy of enhancing Western models rather than building new ones which would challenge China’s disruptive narrative. This may mirror DeepSeek, another Chinese AI that stunned observers at the end of last year by rivaling OpenAI, possibly by leveraging synthetic data from Western sources.

But unlike DeepSeek’s apparent focus on raw computational scale, Manus’s 29-tool augmentation and web browser integration suggest a more practical, execution-driven approach. Yet, the obfuscated browser code and reported jailbreak risks raise security concerns absent in DeepSeek’s profile, hinting at trade-offs for functionality. [tr:trw].