Manus Review (2026): Butterfly Effect's Autonomous AI Agent

Manus Review (2026): The Autonomous AI Agent, Rebooted

Manus is a general-purpose autonomous AI agent from Butterfly Effect, the company behind the Monica assistant. Unlike a chatbot that answers in a single turn, Manus takes a goal, breaks it into steps, and executes them end to end inside its own cloud browser and sandbox, returning finished deliverables such as reports, spreadsheets, slide decks, or working web pages. This hand-reviewed update covers what Manus actually does in 2026, its credit-based pricing, where it shines and where it stumbles, the corporate saga around it, and how it compares with ChatGPT's agent mode and Devin.

What is Manus?

Manus launched as an invite-only preview in March 2025 and went viral almost immediately, positioned as one of the first genuinely "autonomous" general agents rather than a coding-only tool. You describe an outcome in natural language, and Manus plans an approach, then works through it on its own: it opens a virtual browser, navigates websites, runs code in a sandbox, reads and writes files, and calls tools, narrating its steps as it goes. When it finishes, it hands back the actual artifact rather than instructions for you to carry out.

The company behind it, Butterfly Effect, was founded by Xiao Hong and first became known for Monica, a browser-extension AI assistant that reportedly passed ten million users and profitability before Manus existed. Manus is the more ambitious follow-up: an agent aimed at analyst- and operator-style work across research, data, and content, not just software engineering. The core idea is delegation over conversation, closer to handing work to a capable contractor than chatting with a bot.

Manus at a glance

How Manus works

Autonomous cloud sessions. Each task runs on Manus's own infrastructure, with a headless browser, a code sandbox, and a file system. Manus drafts a plan, executes each step, reacts to what it finds, and assembles a final deliverable. Because it runs in the cloud, you can start a task and close your laptop while it works.

Multiple modes. In 2026 Manus offers an Agent mode for full autonomous execution and a lighter Chat mode for quick back-and-forth, along with tiered models (a faster "Lite" tier and heavier "Max"-style tiers on paid plans). Heavier models produce stronger results on hard tasks but consume credits faster.

Concurrent and scheduled tasks. Paid tiers allow multiple tasks to run at once and support scheduled runs, so recurring research or monitoring jobs can execute on a timetable.

Deliverable-first output. Manus is built to return artifacts: a formatted report, a populated spreadsheet, a slide deck, a small website, or a data analysis, rather than a wall of chat text.

Pricing

Manus uses a credit-based model rather than flat unlimited access, and its plan names and allowances have been revised more than once, so confirm current figures before committing. Pricing below was verified in August 2026 and is approximate.

Free plan. $0, with roughly 1,000 one-time starter credits plus about 300 refresh credits per day, one concurrent task, and access to the lighter model tier in Agent mode. Enough to trial the workflow but quickly exhausted on any substantial job.

Entry paid tier. Around $20/month for roughly 4,000 monthly credits, unlocking Chat mode and the fuller model lineup, with more concurrent tasks. This is the practical starting point for regular use.

Mid tier. Around $40/month for roughly 8,000 monthly credits, aimed at heavier individual users.

Top individual tier. Around $200/month for roughly 40,000 monthly credits, for people running frequent or large autonomous jobs. Team plans start from around $20 per seat/month with shared workspace features.

The important caveat is that credits, not dollars, are the real currency, and they burn fast. Complex autonomous runs reportedly consume several hundred credits each, with some multi-step jobs cited in the 500-to-900-credit range, so a single ambitious task can eat a meaningful slice of a monthly allowance. Monthly plan credits generally do not roll over, and daily refresh credits reset without accumulating, so the effective cost of serious use is higher than the headline price suggests.

What Manus does well

Manus is genuinely impressive on open-ended, multi-step knowledge work that would otherwise mean an hour of manual clicking and copying. It is strong at hosted research: gathering information across many web pages, synthesizing it, and returning a structured report. It handles data tasks such as pulling numbers into a spreadsheet and producing basic analysis, and it can generate content deliverables like slide decks and simple web pages from a prompt.

The autonomy is the headline. On tasks it can handle, Manus covers the full loop, from interpreting the request through browsing, computing, and assembling the finished output, without step-by-step supervision, and long jobs run in the background. On the GAIA agent benchmark, Manus reported state-of-the-art results across all three difficulty levels, ahead of comparable deep-research systems, though those figures are self-reported and a 2026 academic audit noted GAIA can be gamed without truly solving tasks, so treat vendor scores as marketing rather than proof.

Capabilities and limits

Manus is powerful but uneven, and its reliability does not yet match its ambition.

The clearest limitation is steering. Once a task is running, there is little room to course-correct mid-execution: if Manus misreads the goal, you often either let it finish and waste the credits or cancel and start over. It also struggles with complex branching logic and can be unreliable on long-running jobs, sometimes stalling, looping, or drifting off-target on the exact kind of ambitious task that motivated using an agent in the first place. A fair summary is high autonomy with medium reliability.

Cost compounds the problem, because credits are consumed whether or not the result is usable, so failed or misdirected runs are not free. For coding specifically, Manus can write and run scripts, but it works in a disposable sandbox without the IDE integration, pull-request workflow, or test-suite support that a dedicated engineering agent like Devin provides. The realistic mental model is a fast, ambitious generalist that needs verification, not an autopilot you can trust unattended.

Pros and cons

Pros

• Genuinely autonomous, end-to-end execution of multi-step tasks
• Deliverable-first output: reports, spreadsheets, slide decks, simple sites
• Cloud sandbox plus browser handles research and light data work well
• Concurrent and scheduled tasks for background and recurring jobs
• Free tier and a ~$20 entry plan make it easy to try

Cons

• Credit consumption is heavy and hard to predict; ambitious tasks are expensive
• Little ability to steer or course-correct mid-execution
• Uneven reliability on complex logic and long-running tasks
• Failed runs still cost credits; monthly credits do not roll over
• Weaker than dedicated agents for serious software engineering
• Ongoing corporate and regulatory upheaval adds uncertainty

Who is Manus for?

Manus fits people with open-ended, repeatable knowledge work they would like to offload: market and competitive research, data gathering and light analysis, first-draft reports and decks, and recurring monitoring jobs. Analysts, founders, marketers, and operators who value a finished artifact over a chat transcript get the most from it, especially those comfortable verifying output and budgeting for credits. It is a weaker fit for developers wanting a reliable coding agent, for anyone who needs tight control over each step, and for cost-sensitive users who cannot tolerate variable, sometimes steep, credit bills.

Manus vs ChatGPT agents and Devin

These tools sit at different points on the agent spectrum. ChatGPT added an agent mode that browses, runs code, and completes tasks inside a hosted environment, and its biggest advantage is distribution: it lives where hundreds of millions of people already work, with no separate app. Manus is a more single-minded autonomous agent, often praised for going further unattended on a task but at higher and less predictable credit cost. For a direct breakdown, see ChatGPT vs Manus.

Devin is a different animal: a cloud agent specialized for software engineering, with repository access, pull-request workflows, and test integration. If the job is code, Devin is the more capable and reliable choice; if it is broad research, data, and content, Manus is the stronger generalist. In practice, many users pair a general agent like Manus for analyst-style work with a specialist like Devin for engineering.

Company, funding, and the Meta saga

Manus's corporate story is unusually turbulent. Butterfly Effect raised roughly $75 million in April 2025 in a round reportedly led by Benchmark, valuing the company near $500 million. Around the same time, its founders relocated from China to Singapore and, by mid-2025, the company reportedly wound down its China-based team to reduce geopolitical risk.

Then the whiplash: Meta reportedly moved to acquire Manus in a deal announced in late December 2025 valued at around $2 billion, with some reports citing figures as high as $3 billion. That did not stick. In 2026, Chinese regulators (reportedly the NDRC) ordered the acquisition unwound on national security grounds, and by August 2026 Manus said it would return to operating as an independent company, with some users asked to back up data created during the Meta period. All of these figures and events are reported by third parties and are best treated as approximate; the practical takeaway is that Manus re-enters late 2026 independent, but under a cloud of regulatory and ownership uncertainty that buyers should factor in.

Verdict

Manus in 2026 is one of the more genuinely capable general-purpose autonomous agents available, and on open-ended research, data, and content tasks it can deliver finished work with striking independence. The trade-offs are equally real: credit consumption is heavy and unpredictable, mid-task steering is limited, reliability wobbles on complex or long jobs, and the company sits in the middle of an unresolved regulatory drama. For users with repeatable knowledge work, a tolerance for verification, and a budget for credits, Manus can save real time, provided expectations, and spending, are set with eyes open. Anyone whose primary need is dependable software engineering, tight step-by-step control, or predictable costs should look at more specialized tools first.

FAQ

Is Manus fully autonomous? Largely, yes, within a task. It plans and executes multi-step jobs on its own in a cloud browser and sandbox, but you have little ability to steer once a run starts, and its output should be verified rather than trusted blindly.

How much does Manus cost? There is a free tier with limited daily credits, an entry plan around $20/month, a mid plan around $40/month, and a top individual plan around $200/month, plus per-seat Team pricing. Everything is metered in credits, and complex tasks can consume several hundred credits each, so real costs depend heavily on usage.

How does Manus compare to ChatGPT's agent or Devin? ChatGPT's agent mode is convenient and built into a tool most people already use; Manus often goes further unattended but costs more in credits; Devin is the stronger choice for actual software engineering. See ChatGPT vs Manus for a direct comparison.

Who owns Manus now? Butterfly Effect. Meta reportedly acquired it in late 2025, but Chinese regulators reportedly ordered that deal unwound in 2026, and Manus has said it is returning to independent operation. Ownership and regulatory details remain fluid, so verify the current situation before making commitments.