Is Manus’s $10,000 Invite a Tech Revolution or a Patchwork AI Hype? In‑Depth Review

Interface design

Manus splits the screen into a chat pane on the left and a sandboxed command‑line pane on the right. After a user query the large language model (LLM) generates a step‑by‑step plan, and each step is displayed and executed in the command‑line window, giving the impression that the AI is directly operating the computer. The design relies on simple, widely available sandbox tools.

Task 1: Novel writing

Prompt: a 5,000‑word biographical novel about the evolution of containers and virtual machines. Manus first calls an LLM (identified as Qwen or Claude 3.7) to create an outline, then iteratively invokes the model for each outline item and stitches the results together. Function‑calling with a Google‑search plugin is used to fetch historical facts. The final text shows inconsistent coherence between chapters, revealing the limitation of current LLMs in long‑text generation.

Task 2: Price‑comparison report

Prompt: compare prices for a Xiaomi AI glasses across Taobao, Xianyu, 1688, and Pinduoduo. Manus builds a plan, attempts to log into each site, and fails to handle authentication. When a page cannot be opened it skips it, sometimes extracting a price from a screenshot without verification, leading to inaccurate results (e.g., reporting a price of 339 CNY without opening the page).

Task 3: Financial analysis and web programming

Prompt: create a detailed Excel valuation model for NVIDIA using data from the Yahoo Finance API. Manus generates Python code to call the API, but the returned stock data are hard‑coded hallucinations, matching only about 30 % of the real values. Errors accumulate across multiple steps, degrading the final report. For a request to build a Google‑style simulation game with NestJS, Manus produces code that fails to run, reports a deployment error, and suggests switching to a single‑page application without completing the deployment.

Core architecture

大语言模型：Qwen，Claude 3.7
搜索工具
虚拟机和自动工具界面
编程工具
静态页面部署工具
内置提示词

All components are mature, off‑the‑shelf solutions. Token consumption is extremely high; a single query can cost roughly 50 CNY at OpenAI pricing.

Open‑source alternatives

OpenManus : high‑star open‑source replacement built in three hours, runs locally without an invite. URL: https://github.com/mannaandpoem/OpenManus AutoMate : AI + RPA tool that turns natural‑language tasks into automated desktop actions. URL: https://github.com/All-Hands-AI/OpenHands OpenHands : multi‑agent “AI programmer” capable of code modification, command execution, web browsing, and API calls. URL:

https://github.com/All-Hands-AI/OpenHands

Conclusion

Manus orchestrates existing LLM calls, sandbox execution, and prompt engineering without introducing new techniques. Its long‑text generation, web‑scraping with authentication, data‑analysis, and code‑generation capabilities are comparable to current LLM‑based pipelines and suffer from hallucination and token‑cost issues.