Manus AI Agent — Hands-On Testing and Pricing of an Autonomous Task-Completing AI Agent
Manus launched in March 2025 billing itself as "the world's first general AI agent." Its biggest differentiator isn't model capability but the fact that it runs
Manus launched in March 2025 billing itself as "the world's first general AI agent." Its biggest differentiator isn't model capability but the fact that it runs tasks by default on an asynchronous cloud virtual machine — close your browser and it keeps going. After testing 12 tasks, the four categories it genuinely delivers on reliably are data gathering, spreadsheet compilation, slide decks, and website generation; tasks requiring account logins, payments, or cross-system writes have a markedly higher failure rate. Pricing has evolved from the hotly disputed US$39 entry tier of its early days into a Credit-based system, where cost depends heavily on task complexity rather than hours used. What Manus Is: The Structural Difference from ChatGPT and Claude Manus is an autonomous AI agent built by the Chinese team Monica (Butterfly Effect), released in preview on 6 March 2025, with the company later relocating its headquarters to Singapore. It doesn't train its own foundation models; instead it layers a planner, a sandboxed virtual machine, and a toolset on top of existing models such as Anthropic's Claude. That structural difference determines where it fits. The chat interfaces of ChatGPT and Claude are essentially synchronous and turn-based: you ask, they answer, and long tasks advance only as you keep prompting. Once Manus receives an instruction, it first produces a todo.md task list, then works through it item by item inside a cloud Linux virtual machine — opening a browser, downloading files, writing Python scripts, running shell commands, producing output files — all without you present. The benchmark figures published at launch are what kicked off the attention: "Manus achieved then-SOTA scores across all three difficulty levels of the GAIA benchmark, reaching 86.5% on Level 1 (source: GAIA Benchmark Leaderboard / Hugging Face)" . GAIA is a general-assistant evaluation jointly proposed by Meta AI, Hugging Face, and the AutoGPT team; its questions demand multi-st
FAQ
What Manus Is: The Structural Difference from ChatGPT and Claude
Manus is an autonomous AI agent built by the Chinese team Monica (Butterfly Effect), released in preview on 6 March 2025, with the company later relocating its headquarters to Singapore. It doesn't train its own foundation models; instead it layers a planner, a sandboxed virtual machine, and a toolset on top of existing models such as Anthropic's Claude. That structural difference determines where it fits. The chat interfaces of ChatGPT and Claude are essentially synchronous and turn-based: you
Hands-On: What Actually Gets Finished, and What Stalls
Across 12 everyday office and research tasks, deliverability splits into clear groups. The breakdown below sorts them into "passed first time," "needed human intervention," and "couldn't complete." Task Types It Delivers Reliably Multi-source data gathering and comparison tables : Given "compare the pricing, storage, and regional restrictions of five cloud drive services and output an Excel file," Manus actually opens each official site, scrapes the pricing pages, and produces an .xlsx . Average
Pricing Structure: How Credits Work, and What It Really Costs
Manus currently uses a Credit system rather than billing by time or by number of conversations. Credits consumed per task depend on the number of execution steps, tool calls, and virtual machine runtime — meaning that under the same subscription tier, the more complex your tasks, the fewer of them you get. The official tiers break down roughly as follows (defer to the official pricing page; plan contents change frequently): Free : A small daily Credit allowance, enough only to test a single shor
Who It Suits, and Who It Doesn't
Manus's value concentrates in tasks that are repetitive, time-consuming, but quickly verifiable. Users meeting the following conditions see the highest return: You regularly do market or competitor data gathering, mostly from public web sources. You often need to turn unstructured data (PDFs, web pages, messy CSVs) into structured tables. You can judge for yourself whether the output is correct — that is, you already understand the subject and simply don't want to spend the time. Conversely, the
Related Guidebooks
Reviewed and verified by FeiYueh · Last verified 2026-09-17. Independently maintained — not AI-generated boilerplate.
← Back to Blog