In just a week, over 2 million users have joined the waiting list for China’s Manus general AI agent. Touted as the “second DeepSeek moment” for China, Manus AI is currently in closed beta and accessible only via invitation.
The excitement has surged, with many hailing Manus AI as a major breakthrough and China’s answer to OpenAI’s Deep Research agent especially as the country continues to push AI innovation at lower costs. However, the hype is largely overblown, fueled in part by AI influencers making bold claims on social media. Here’s why I believe Manus AI is a promising development but not a true breakthrough.
Why Manus AI Falls Short of a Breakthrough
DeepSeek was a breakthrough because it successfully replicated OpenAI’s RL method, delivering performance comparable to O-series reasoning models all while operating on a fraction of OpenAI’s training budget. The DeepSeek team further cemented its impact by introducing and open-sourcing the GRPO training method, enabling other labs to develop frontier-class reasoning models.
These were genuine innovations, made even more impressive by the fact that DeepSeek achieved them despite U.S.-imposed GPU constraints. In contrast, the Manus general AI agent is not built on groundbreaking advancements. Instead, it combines Anthropic’s Claude 3.5 Sonnet model with several fine-tuned Qwen models and relies on the open-source Browser Use project.
While better integration and tooling offer advantages, the true breakthrough lies in developing frontier-class models specifically optimized for agentic tasks. Anthropic’s Claude 3.5 Sonnet is already one of the most capable AI models for such tasks, including coding. In fact, the Manus team is internally testing the upcoming Claude 3.7 Sonnet unified model and considers it “promising.”
Ultimately, building powerful AI models remains the key differentiator—and will continue to be for the foreseeable future. That said, the Manus AI team deserves credit for effectively chaining multiple tools and environments to accomplish tasks. As mentioned earlier, it’s a promising step toward a more agentic AI future.
Manus AI Agent Falls Short
We haven’t had direct access to Manus AI, but some X users who received early access have shared their experiences. Biomedical scientist Derya Unutmaz tested Manus alongside OpenAI’s Deep Research agent and posted his findings on X.
According to Unutmaz, Deep Research completed the task in 15 minutes, whereas Manus took 50 minutes and still failed to finish. He also noted that Manus does not reference sources the way Deep Research does.
Similarly, X user teortaxesTex tested the Manus agent and found that it excels at regurgitating information like traditional LLMs rather than performing true agentic tasks. Another user, TheXeophon, shared his experience, noting that Manus completely failed to mention the Nintendo Switch while researching the gaming console market.
In fact, the viral video claiming that the Manus AI agent automated 50 tasks turned out to be fake. Yichao “Peak” Ji, Manus’ chief scientist, debunked the video, stating, “This video is definitely NOT Manus,” followed by a laughing emoji.
Despite its early missteps, it’s important to remember that Manus is still in closed beta, and dismissing it entirely would be premature. At the same time, it’s crucial to approach new AI products with a measured perspective. While Manus may not be a groundbreaking innovation, it represents an ambitious step in the right direction.
As AI models continue to improve at agentic tasks, products built on top of them will also evolve. The Manus AI team has already stated that the agent will undergo significant enhancements before a broader public release. Whether it lives up to the hype remains to be seen, but it’s certainly a development worth keeping an eye on.