In just one week, more than 2 million users have joined the waiting list for China’s Manus general AI agent, currently available only through invitation in closed beta. This surge of interest has led many to dub Manus AI as China’s “second DeepSeek moment” and a potential rival to OpenAI’s Deep Research agent. But does this new entrant truly represent a breakthrough in AI development, or is the excitement premature?
Understanding the Hype
The frenzy surrounding Manus AI stems partly from China’s growing reputation for delivering AI innovations at competitive prices. Social media platforms are buzzing with influencers making bold claims about Manus’s capabilities, further fueling public interest. However, a closer examination reveals that while Manus represents a promising development, characterizing it as a breakthrough may be an overstatement.
Manus | A general AI agent that bridges minds and actions – https://manus.im/
China’s Manus AI ‘agent’ could be our 1st glimpse at artificial general intelligence – https://www.livescience.com/technology/artificial-intelligence/chinas-manus-ai-agent-could-be-our-1st-glimpse-at-artificial-general-intelligence
MIT | Everyone in AI is talking about Manus. We put it to the test – https://www.technologyreview.com/2025/03/11/1113133/manus-ai-review/
Why Manus AI Is Not (Yet) a Breakthrough
DeepSeek earned its breakthrough status by successfully replicating OpenAI’s reinforcement learning methods to deliver performance comparable to advanced reasoning models—all on a significantly smaller budget than OpenAI’s training costs. Additionally, DeepSeek’s introduction and open-sourcing of the GRPO (Gradient-based Reinforcement through Policy Optimization) training method provided other labs with tools to train frontier-class reasoning models.
These innovations were particularly impressive considering the GPU constraints imposed on China by U.S. trade restrictions.
In contrast, Manus appears to be an integration of existing technologies rather than a fundamental innovation. The system combines Anthropic’s Claude 3.5 Sonnet model with several fine-tuned Qwen models and relies on the open-source Browser Use project. While this architecture demonstrates skilled integration, it doesn’t represent the same level of pioneering advancement that DeepSeek achieved.
The Manus team is reportedly testing the newer Claude 3.7 Sonnet unified model internally, finding it “promising” for their purposes. This underscores a key point: developing capable foundation AI models remains the primary competitive advantage in the industry, even as application layers become increasingly sophisticated.
OpenAI – Introducing deep research – https://openai.com/index/introducing-deep-research/
Anthropic’s Claude 3.5 Sonnet – https://www.anthropic.com/news/claude-3-5-sonnet
Anthropic’s Claude 3.7 Sonnet – https://www.anthropic.com/claude/sonnet
Early Performance Challenges
Early access reports from users indicate that Manus still faces significant performance challenges. Biomedical scientist Derya Unutmaz shared a comparative test on X (formerly Twitter), noting that while OpenAI’s Deep Research completed his assigned task in 15 minutes, Manus ran for 50 minutes and ultimately failed to complete the same task. Unutmaz also observed that Manus doesn’t reference sources in the same way Deep Research does, potentially limiting its utility for academic or research applications.
Derya Unutmaz – https://x.com/deryatr_
A Promising Direction Despite Limitations
Despite these initial shortcomings, it’s important to recognize that Manus is still in closed beta. The development team has already announced plans for significant improvements before a wider public release.
The Manus AI team deserves credit for their ambitious approach to chaining multiple tools and environments to complete complex tasks. This integration-focused strategy may not represent a breakthrough in fundamental AI capabilities, but it does represent a promising step toward more sophisticated agentic AI systems.
The Future of Agentic AI
As foundation AI models continue to improve at handling agentic tasks, products built on these models will naturally evolve and become more capable. The current limitations of Manus AI may be temporary, overcome through iterative development and integration with more advanced underlying models.
Whether Manus ultimately lives up to its current hype remains uncertain, but it represents a noteworthy development in China’s AI landscape and a significant attempt to create competition in the emerging field of AI agents. For industry observers and potential users, Manus AI warrants continued attention as it moves toward public release and further refinement.
As the global AI race continues to accelerate, systems like Manus remind us that integration and application engineering represent important complementary skills to fundamental model development—both will be crucial for creating truly useful AI systems in the future.
Artificial intelligence (AI) – https://en.wikipedia.org/wiki/Artificial_intelligence
Grok. A Deep Dive into The Evolution of AI Assistants – https://cialu.net/grok-a-deep-dive-into-the-evolution-of-ai-assistants/
5 Strategic Approaches on How to Use ChatGPT to Accelerate Your Learning – https://cialu.net/5-strategic-approaches-on-how-to-use-chatgpt-to-accelerate-your-learning/