What Is Manus AI? The Autonomous Agent, Honestly Assessed (2026)

Manus has had the strangest eighteen months of any product in AI. It arrived in March 2025 as an invite-only demo people were reselling access to for four figures. Meta bought it for a reported $2 billion in December 2025. China’s economic planner blocked the deal in April 2026 and ordered it unwound. As of 1 September 2026 Manus is independent again, with the same founders running it. All of which is more interesting than useful if what you want to know is whether the thing works — so this piece covers what Manus does, what independent testers find when they run it on real work, and what happened to the company and why it might matter to you.

Quick answer

What Manus actually is

The word “agent” gets used for three different things, so it is worth being precise about which one Manus is. It is not a chatbot with a browser bolted on, and it is not a copilot suggesting things inside your editor. When you give Manus a task, it provisions a cloud virtual machine for that task — a real Linux environment with a file system, a terminal, a code interpreter and a Chromium browser — and works inside it. You can watch it click, type and run commands in real time, or close the tab and come back to a finished artefact. The name comes from MIT’s motto Mens et Manus, “mind and hand”, and that is the whole pitch: other systems give you the mind, this one is meant to give you the hand.

The September 2026 feature set is broad: Wide Research (a parallel multi-agent mode for large-scale gathering), Manus Slides (research to finished deck with speaker notes), a website and web-app builder with GitHub export and Stripe integration, scheduled tasks, Projects as persistent workspaces with standing instructions, meeting minutes, data visualisation, and a Browser Operator extension that drives your local browser using your own authenticated sessions rather than the cloud sandbox. It reaches outside tools through MCP servers, including custom ones you host, plus Zapier, Slack and a public API. There are desktop and mobile apps. The current stable release is Manus 1.6, shipped 15 December 2025, in three tiers: Lite, standard and Max.

How it actually works

Manus’s chief scientist Yichao “Peak” Ji published an unusually candid engineering post on the system’s design, and it remains the best primary source on how it behaves — including why it behaves badly in certain conditions.

The headline decision is that Manus deliberately does not train its own end-to-end agentic model. It bets entirely on context engineering on top of other labs’ frontier models: if model progress is the rising tide, Ji writes, Manus wants to be the boat rather than the pillar stuck to the seabed. At launch a co-founder confirmed the stack ran on Anthropic’s Claude Sonnet plus fine-tuned Qwen models, orchestrated by Manus’s own planning and tool-use layer.

The loop is conventional — analyse, plan, execute, observe, append, repeat. The interesting parts are the workarounds for the ways it fails.

• A typical task takes around 50 tool calls. That is Manus’s own published average, and it explains most of the reliability picture below. Fifty sequential opportunities to drift is a lot.
• The file system is the memory. Rather than compressing context when it overflows, Manus reads and writes files in the sandbox, keeping compression restorable — drop a page’s content, keep its URL. Context windows are finite; the sandbox disk is not.
• It recites its own to-do list. The todo.md it creates and ticks off pushes the global plan back into the model’s recent attention span, countering goal drift over a long run.
• It keeps its mistakes in context on purpose. Failed actions and stack traces stay in the trace, because erasing failure removes the evidence the model needs in order not to repeat it. Ji argues error recovery is the clearest indicator of genuine agentic behaviour and the thing public benchmarks least measure — a notably honest thing for a vendor to write.

One number worth internalising: Manus reports an average input-to-output token ratio of roughly 100:1. Agents are overwhelmingly a reading cost, which is why credits drain in ways that feel disproportionate to the output you got.

What it does well

Stripping out both the launch hype and the backlash, a consistent picture emerges from people who run it at volume.

One-shot research and synthesis. Give Manus a broad, open-ended question with a defined deliverable and it goes away, reads widely, and comes back with something structured and usable. This is the task shape it was built for, and where independent testers are most consistently positive.

Document, deck and dataset production. The gap between “here is my analysis” and “here is a formatted deck with speaker notes and charts” is where a lot of knowledge work actually lives, and Manus closes it in one pass. Wide Research covers the breadth cases — process 200 of these, compare 50 of those. It will also get something built that did not exist before: prototypes, scrapers, small web apps, quick data pipelines. Not production code, but a working artefact in minutes rather than an afternoon.

Economics on ad hoc work. A complex task costs a few dollars in credits and finishes in minutes — favourable against a researcher’s hourly rate, provided the task really is one-off. The moment it becomes a repeated process, the calculation inverts, for reasons below.

Where it falls short, with evidence

Start with the launch, because the gap between demo and reality set the tone for everything after. Manus’s marketing led on GAIA, a benchmark for general assistant capability, claiming roughly 86.5% on Level 1, 70.1% on Level 2 and 57.7% on Level 3 — ahead of OpenAI’s Deep Research. The scepticism was immediate and came from serious people. Zvi Mowshowitz’s contemporaneous write-up characterised Manus as a Claude wrapper with a jailbreak and a few dozen tools, optimised for GAIA specifically, behind an influencer-led marketing push; he also noted the launch chart’s x-axis started at 30% rather than zero, flattering the lead. TechCrunch’s assessment was headlined, with some restraint, that Manus probably was not China’s second DeepSeek moment.

The deeper problem is structural rather than specific to Manus. Work on agent benchmarks, including a UC Berkeley RDI review of benchmark trustworthiness, has shown that evaluations such as GAIA can be scored well on without genuinely solving the underlying task. A GAIA number measures capability in controlled conditions, not reliability across a hundred messy real-world tasks — which is what you are actually buying.

Non-determinism is the most-cited complaint. In the most substantial published hands-on assessment — Martin Kelly at Botonomy, who ran 50-plus tasks over three months — the same competitive-analysis prompt run three times in one week returned three meaningfully different outputs: different sources cited, different conclusions, different data points highlighted. For exploratory work that is tolerable. For a process that has to produce the same shape of answer every Monday, it is disqualifying. His summary: Manus excels at one-shot research and document generation, and struggles with multi-session workflows and deterministic repeatability.

Long chains degrade. Community-reported figures — not vendor data, not peer-reviewed, worth treating as directional — put success on open-ended research at roughly 70 to 80%, with a sharp fall-off past about five dependent steps. The most commonly reported failure modes are hallucinated clicks, browser timeouts and anti-bot blocks. That last is a structural constraint on every cloud-sandbox agent: a datacentre IP driving a headless browser looks exactly like what bot detection exists to stop.

Confident fabrication. The documented cases are the ordinary kind, such as a user reporting that Manus produced a company revenue figure during market research that appeared in none of the sources it cited. Not a Manus-specific flaw, but autonomy multiplies its cost: a chatbot’s hallucination sits in front of you, whereas an agent’s is carried forward into fifty subsequent steps and into the finished deliverable.

Manus now acknowledges this in-product. Since June 2026 it displays a “high-impact topic” notice on conversations touching health, finances or legal rights, advising independent verification and telling users to treat Manus as “a starting point, not a final authority”. It blocks nothing. As vendor disclosures go, it is a reasonable one.

One criticism has aged out: earlier reviews complained Manus had no native integrations and could not push data into a CRM or pull from analytics. MCP connectors, custom MCP servers, Slack, Zapier and a public API have since closed most of that gap. Check that claim before relying on any review written in early 2026.

Pricing and credits

Everything here comes from Manus’s own help centre, last updated 16 March 2026, with the vendor’s own caveat that the live pricing page is authoritative.

• Free — $0/month. Chat Mode, plus Manus 1.6 Lite in Agent Mode only. 300 refresh credits daily. One concurrent task, two scheduled tasks. You do not get 1.6 or 1.6 Max on this tier.
• Pro, from $20/month — from 4,000 credits a month. All three model tiers in Agent Mode, advanced research, website deployment, slide generation, Wide Research, 20 concurrent and 20 scheduled tasks, early access to beta features.
• Pro, from $40/month — from 8,000 credits a month, same feature set, with a 7-day free trial.
• Team, from $20 per seat/month — everything in Pro, plus SSO, data training opt-out, team usage analytics, internal access control and shared slide templates.
• Annual billing saves 17% across the paid tiers.

If you see Basic and Plus in older write-ups, both were renamed Pro; existing subscribers kept their original prices on auto-renewal.

On credits: consumption scales with task complexity, not time or output length. The dashboard estimates cost before a task starts. Plan credits reset monthly; separately purchased add-on packs never expire; Team and Enterprise use a shared pool.

The practical warning is that burn is unpredictable, and the 100:1 input-to-output ratio is why. Two superficially similar research tasks can cost very different amounts depending on how many retries and dead ends the agent hits internally — and per Manus’s own design philosophy, those dead ends are deliberately kept in context rather than discarded. Manus’s published advice for reducing burn is sound: be specific rather than broad, use Chat Mode for anything that does not need the full agent, batch related work into one task, and review intermediate output before letting a long run continue.

Data, ownership and governance

This is the section people search for, so here are the facts without editorialising in either direction.

Who owns it. Butterfly Effect was founded in China in 2022 by Xiao Hong, joined by Ji Yichao (chief scientist) and Zhang Tao (chief product officer). Manus launched in invite-only beta on 6 March 2025 and went viral; codes were resold on Chinese secondary markets, with China Daily reporting ¥50,000 to ¥100,000 and the Wall Street Journal reporting resales above $1,000. Backers include ZhenFund, Tencent and HSG, plus a roughly $75 million Series B led by Benchmark in April 2025 at about a $500 million valuation.

The relocation. After that round the company moved its headquarters from Beijing and Wuhan to Singapore, the three co-founders relocating in mid-2025, with a separate Singapore entity taking over the product outside China. Of roughly 120 China-based staff, about 40 core technical people moved and the rest were laid off. Manus closed its Chinese-language social accounts, blocked mainland access, and shelved a planned Chinese version announced with Alibaba’s Qwen team. Context: the US Treasury had opened a review of Benchmark’s investment under the Outbound Investment Security Program in May 2025, and chip export controls were tightening. That review largely faded after the move.

The Meta episode. Meta announced in December 2025 that it would acquire Manus, reportedly for more than $2 billion. China’s Ministry of Commerce opened an evaluative investigation in January 2026, exit bans were issued for executives under scrutiny in March, and on 27 April the National Development and Reform Commission blocked the deal outright. Meta split operations and data in June and cut ties. Manus announced its return to independence on 11 August and formally resumed independent operations on 1 September, founding team intact; some users had to back up and restore data through a portal that remains open with no deadline. As of mid-September it is reported to be raising around $500 million at a roughly $4 billion valuation — press reporting, not a company announcement.

What that means for your data. Manus is Singapore-domiciled and holds a SOC 2 Type 2 attestation published through its trust centre. Its privacy documentation describes technical, organisational and physical safeguards, with the standard caveat that no online service is completely secure. Retention is “as long as necessary”, with longer holds permitted for legal, accounting, fraud-prevention and claims purposes.

Two specifics matter before you put company data through it. Data training opt-out is listed as a Team plan feature, not a Pro one, so policy rather than usage may dictate which tier you buy. And on Team plans a Team Owner can access all session data for team members, including task steps and content, and can export it; Administrators see a subset. Individual accounts are private by default, and Team retention is set by the Owner or Administrator rather than the user.

The governance point the past year established rather than theorised: a Chinese-founded company can relocate its headquarters, staff and legal domicile, and Chinese regulators can still block what happens to it. Whether that is relevant depends entirely on what you put in the sandbox. For public-source market research it is a non-issue. For client lists, unreleased financials or regulated data, it is a question for your own vendor process — the same one you would ask of anyone, just with a more concrete example to reason from.

Manus vs ChatGPT agents vs Devin

These three get compared constantly and are not really substitutes.

Manus vs ChatGPT. ChatGPT is a general assistant that has an agent mode; Manus is an agent that has a chat mode. That inversion explains most of the practical differences. ChatGPT is better if you want one subscription covering conversation, writing, images, coding help and occasional autonomous work, and it has the deeper enterprise governance story and wider connector ecosystem. Manus goes further unattended, produces more finished artefacts — decks, sites, datasets — and gives you a real Linux box with a shell rather than a constrained browsing environment. For what OpenAI’s version actually does, see our breakdown of ChatGPT agent mode; for plans and model tiers, our ChatGPT hub. Both entry tiers are $20 a month, so this is a question of task shape rather than budget.

Manus vs Devin. Devin is narrowly a software engineer, working against repositories, pull requests and tickets, and priced as infrastructure: $20 a month on Core plus $2.25 per Agentic Compute Unit, where one ACU is roughly 15 minutes of active work, or $500 a month on Team for 250 ACUs at $2.00 each with unlimited concurrency. If the task is “implement this ticket in this codebase”, Devin is right and Manus is not — Manus’s code generation is widely described as hit-or-miss. If the task is “research this market and give me a deck”, the reverse holds. Our Devin AI page covers that side in detail.

The useful model: Devin is specialised and deep, Manus general and broad, ChatGPT general and shallow-but-safe. Pick by task shape, not by which benchmarked highest.

So is Manus AI worth it?

Yes, if your work involves a steady stream of one-off research questions, competitive scans, briefing documents, decks or quick data analyses — and you read the output before anyone else does. At $20 a month against the cost of your own hours, the case makes itself. Start on the free tier; 300 daily credits is enough to find out whether it fits your task shape.

No, if you need the same process to produce the same shape of result every time, if the work involves write access to live systems, or if wrong-but-confident output is expensive in your context. Those are the exact conditions under which the evidence says Manus degrades, and no amount of prompt tuning fixes a non-deterministic system.

With care if you are putting sensitive or regulated data in it. Buy Team for the training opt-out, understand that your Team Owner can read and export everything you run, and make the ownership question an explicit part of your assessment rather than something you discover later.

The fairest summary is that Manus was oversold at launch and has been underrated since. It is not the general-purpose autonomous worker the March 2025 demo implied, and the people who said so at the time were right. It is also a genuinely capable research and production agent a lot of people get real value from daily. Both things are true.