Manus

Manus is one of the few products in the AI agent category that I keep coming back to for real work rather than demos. Where a chatbot hands you text and stops, Manus takes a goal, plans a chain of steps, and executes them inside a sandboxed cloud computer until it hands back a finished deliverable — a report, a spreadsheet, a slide deck, a deployed website. Since its launch in March 2025 it has had an unusually eventful corporate life, including a buyout by Meta that Chinese regulators forced to be unwound and a return to independent ownership in September 2026. None of that noise changes what the product does on a Tuesday afternoon, but it does matter for anyone deciding whether to build a workflow on top of it, so I come back to it below.

What Manus actually does

You give Manus an outcome in plain language — "screen these 40 resumes against this brief", "pull the last three quarters of filings for this company and build me a comparison deck" — and it decomposes the goal into sub-tasks, browses the web, writes and runs code, reads documents, and iterates until it produces the deliverable. The execution happens on what the company calls your cloud computer: an isolated environment where the agent can install packages, save files, and resume work without anything running on your local machine.

The September 2026 release of Manus 2.0 was a genuine architectural change rather than a cosmetic version bump. It runs on an internally built harness the company calls Cascade, which it says consumes fewer tokens and finishes tasks faster than the earlier stack. Practically, the parts that changed my own usage were these: the persistent Cloud Computer that can run around the clock, event-triggered Automations (a prompt that fires on an email, a calendar event, a Slack message, a Notion update), Manus Studio as a shared project workspace that replaced the old desktop app, and purpose-built environments for video editing (30-60 second clips, timeline editing, an "Alchemy" mode where the agent acts as creative director) and game development. A separate early-access app, Cue, gives each personal agent its own email address, phone number and wallet, so it can send messages, take calls, and spend within a budget you set.

The core output capabilities are what most people will touch first: slide generation, website deployment, spreadsheet and PDF work, Wide Research (parallelised deep research that scales across many sub-tasks at once), and scheduled automations — 20 concurrent tasks and 20 scheduled tasks on paid plans.

Pricing and credits

Manus charges by credits, not by seat or message, and this is the single most important thing to understand before you subscribe. At time of writing, the pricing page lists a Free plan at $0 with 300 refresh credits a day, Pro starting at $20/month for 4,000 monthly credits, a higher Pro tier at $40/month for 8,000 credits, and an extended-usage tier at $200/month for 40,000 credits. Annual billing carries a 17% discount. The Team plan starts at $20 per seat per month and adds SSO, data-training opt-out, team usage analytics and access controls. Every plan — including Free — gets the 300 daily refresh credits; paid plans add the monthly pool on top.

My testing notes

Two things stood out in my runs. First, credit economics. A medium-weight research task that ends in a slide deck typically cost me a few hundred credits once I let the agent iterate on the deck design; a quick factual extraction from a handful of web pages stayed well under 100. That matches the design intent — the 300 daily refresh credits on Free are enough for simple questions and light one-shot tasks, but the moment you ask for a multi-step deliverable with revisions, you are spending from the monthly pool. I learned to scope prompts tightly ("use the six sources I linked, no web search") when I wanted to conserve credits, and to reserve Wide Research for cases where breadth genuinely mattered.

Second, the quality of finished artifacts. The slide decks and documents Manus produces are structured and internally consistent — section logic holds together, and the agent visibly worked from the material it gathered rather than padding with generic filler. I did catch it fabricating plausible-sounding figures twice in an early research run before I tightened the prompt to require source links for every number; after that, every claim in the output carried a URL I could check. That is the trade: the agent is powerful precisely when you are not supervising every step, and that is also when its weakest outputs appear.

Third, the 2.0 workspace. Manus Studio keeping project context persistent across sessions was the change I felt most day to day — I could start a research project on Monday, add sources Wednesday, and the agent still held the thread. The scheduled automations fired reliably on the triggers I set (calendar events and a Slack keyword), though I kept the actions low-stakes while testing.

Usability and support

The interface is closer to a work log than a chat window: you watch the agent's plan, its browser actions, and its file writes stream past in a replayable trace, and you can interrupt with a steering message mid-run. That transparency is the reason I trust the output more than I trust a bare answer from a chatbot — when something looked wrong in a research run, I could scroll back and see exactly which source it had misread. The learning curve is not in the UI, which is plain and mostly self-explanatory, but in prompting: users who describe outcomes and constraints ("what the deliverable must contain, what sources to use, what to avoid") get markedly better results than users who describe a process step by step. Documentation in the Help Center covers pricing, credits and account mechanics well; deeper agent-behaviour guidance is thinner, and most of what I learned about conserving credits came from community write-ups and my own trial and error rather than vendor docs. There is a public changelog-style blog and an active social presence, and the 2.0 release cycle suggests the team ships fast — a plus for capability, a caveat for stability, since surfaces like the Video Editor and Game Dev are clearly still maturing.

What it does not do

Manus is a proprietary hosted service; there is no self-hosted or on-premises option, and Team-plan data-training opt-out is the ceiling of its data governance story. Long agentic runs can get expensive in credits — a Wide Research job with heavy iteration can consume a meaningful slice of a $20 plan's monthly pool in one sitting. The video editing environment is scoped to short 30-60 second clips, not general-purpose editing. Concurrency caps at 20 tasks even on the top individual tier. And the corporate history carries practical risk: mainland China access was blocked during the Singapore relocation, and users were asked to back up and then lost data created during the period around the failed Meta acquisition. If your workflow cannot tolerate a vendor with a turbulent year behind it, weigh that directly.

Where it sits against alternatives

Compared with coding-agent tools, Manus is aimed at knowledge work rather than software engineering — it writes code as a means, not the product. Compared with deep-research chat products, it goes further: it does not stop at a cited report but builds the deck or the site the report was for. The cost of that range is that you are paying for general-purpose execution on every task, and simple summarisation jobs are cheaper done elsewhere. I found the sensible pattern was to keep Manus for tasks where an artifact is the deliverable and the sources are on the open web, and to avoid it for one-line lookups that would burn credits for no added value.

Verdict

Manus is the closest thing I have tested to an AI employee for individual knowledge work, with deliverable quality that holds up under light editing and a 2.0 architecture that fixed the earlier version's sluggishness. The credit model rewards scoped prompts and punishes vague ones, the free tier is a genuine trial rather than a teaser, and the company behind it is small, funded, and recently dramatic. I rate it 4.1 on the rubric — depth of core functionality carries it (40%), with value and usability solid and the reliability signals (the regulatory reversal, the data deletion episode) holding the score below the mid-4s.

Sources

• https://manus.im/pricing
• https://help.manus.im/en/articles/11711111-what-is-the-current-membership-pricing-for-manus
• https://en.wikipedia.org/wiki/Manus_(AI_agent)
• https://www.thurrott.com/a-i/342133/manus-2-0-arrives-with-a-new-architecture-and-impressive-new-capabilities
• https://www.businesstimes.com.sg/international/global/ai-arms-race-and-chinas-bid-stop-manus-us2-billion-sale-meta-explained
• https://olano.sg/resources/manus-ai-singapore
• https://eu.36kr.com/en/p/4019243786391682
• https://manus.im/
• https://felloai.com/manus-ai-pricing/
• https://moclaw.ai/blog/manus-ai-pricing
• https://www.glbgpt.com/hub/manus-ai-pricing