Manus AI Review 2026: Manus 2.0, Cue, and What Credits Really Cost

Manus AI is the autonomous agent that does not ask permission. Where most AI assistants hand you a draft and wait, Manus spins up a full cloud computer — browser, file system, terminal, connected apps — and works the task to completion while you do something else.

That premise made it viral in March 2025. Eighteen months later it is a real product with real revenue, a blocked $2 billion acquisition behind it, and as of September 28, 2026, an architectural rewrite called Manus 2.0.

This review covers what Manus 2.0 actually changed, how the credit system bills you, where the agent genuinely fails, and who should pay for it.

What Is Manus AI?

Manus is a general-purpose autonomous AI agent built by Butterfly Effect Pte. Ltd., a company founded in 2022 by Xiao Hong, with Ji Yichao as chief scientist. It launched in invitation-only beta on March 6, 2025.

The architecture is multi-agent. A central controller decomposes your goal into sub-tasks and delegates them to specialized sub-agents — web browsing, code execution, data analysis, file management — each running inside a sandboxed virtual machine in Manus's cloud.

That sandbox is the whole point. Because the agent operates on its own machine rather than yours, it can download files, run scripts, install dependencies, and click through multi-step web flows without any risk to your local environment. You close the laptop; the task keeps running.

Butterfly Effect relocated from Beijing and Wuhan to Singapore in mid-2025 following a $75 million Series B that valued the company at $500 million. By December 2025 its revenue run rate had reached $125 million, up from $90 million in August.

What Changed in Manus 2.0

Manus framed the September 28, 2026 release as "not a version update" but a new architecture with new products. The substance backs that up more than most vendor launch language does.

Cascade: The New Agent Harness

Cascade replaces the execution harness underneath Manus. Manus's internal benchmarks for one configuration report 23.2% fewer tokens, 28.2% less time to completion, and 32% lower run cost versus the prior system.

Treat those as vendor-reported. They were published without independent verification at launch, and "one benchmark configuration" is doing quiet work in that sentence. But the direction matters: on a credit-metered product, a 32% cost reduction is a pricing change disguised as an engineering release.

Studio Replaces the Desktop App

Studio is a project-based collaborative workspace that supersedes the old Manus desktop app. It folds editing directly into the agent environment:

• Document, spreadsheet, slide, and code editing
• A timeline video editor with editable layers for clips, images, text, and audio
• Alchemy mode, combining video and code generation
• Game Dev, supporting playable projects, with a "Max Ultra" tier for production quality
• An edit panel for live preview and selective changes
• Remote Control, authorizing desktop operations from your phone

The shift here is from "agent that returns a file" to "workspace where you and the agent revise the file together." That directly addresses the most common complaint about autonomous agents: that a 90%-correct deliverable you cannot edit in place is worse than a 70%-correct one you can.

Cue: Agents With Their Own Email and Wallet

Cue is a standalone app at cue.im for creating and deploying personal agents. Each agent gets its own email address, phone number, digital wallet, and computer.

That combination is unusual. A Cue agent can send messages, take phone calls and hand you a summary, delegate to other agents in a group chat, and transact within a preset budget — ordering from local services, for example. Early access is free with an invite code, limited to the first 1,000 users, on web and desktop, with iOS pending App Store review.

Giving an autonomous agent a wallet and a phone number is either the obvious next step or the part everyone regrets, depending on how the budget guardrails hold up. It is too new to judge.

Cloud Computer and Automations

Cloud Computer is a dedicated 24/7 environment for work that needs persistence — always-on automations, multiplayer game servers, anything that must stay reachable when your device sleeps. It is a separate paid add-on.

Automations let workflows trigger on events rather than schedules: a new email, an ad performance shift, a calendar event, a Slack message, a Notion update. You configure them in natural language.

Automations is the more consequential feature. Scheduled tasks made Manus a cron job; event triggers make it something closer to a colleague who notices things.

Wide Research

Wide Research is Manus's flagship research mode, and the clearest demonstration of the parallel sub-agent architecture. Instead of one agent reading sources sequentially, it fans out many sub-agents across sources simultaneously.

Manus has cited GAIA benchmark results of 86.5% pass@1 on Level-1 tasks and 57.7% on Level-3, above the comparison figures it published for competing deep-research systems.

Read those claims carefully. Manus has not released GAIA or HLE benchmarks for Wide Research specifically, nor direct head-to-head comparisons, so the Wide Research claims remain largely qualitative. Independent reviewers have also noted that academic benchmark scores translate unevenly to real research tasks, and that Manus has sometimes missed key details in practice.

It is also the most expensive thing you can run. Wide Research consumes 800–1,500+ credits per run, and heavier runs have been reported at around 4,000 credits — an entire month's Pro allocation in one command.

Pricing: The Credit System Is the Catch

Manus bills a subscription plus a pool of credits consumed by actual work.

Annual billing saves roughly 17%. At both $20/4,000 and $40/8,000, a credit costs $0.005 — the tiers buy volume, not a better rate.

Credit consumption scales with three things: LLM tokens for planning and output, virtual machine time for browser and file work, and any third-party APIs the task calls. Reported rates run roughly 11–14 credits per active minute of execution.

Typical costs:

• Simple query or web lookup: 10–50 credits
• Web research task: 100–300 credits
• Code generation and execution: 200–500 credits
• Deep research: 500–900 credits
• Web app build: 500–1,000+ credits
• Wide Research: 800–1,500+ credits

Manus publishes three worked examples: data analysis and visualization at 15 minutes for 200 credits; website design and deployment at 25 minutes for 360 credits; app development and integration at 80 minutes for 900 credits.

Here is the real problem: Manus does not estimate a task's credit cost before it runs. You launch a job without knowing what it will charge you. On a $20 plan with 4,000 credits, four or five ambitious tasks can exhaust the month. Cascade's claimed 32% cost reduction directly targets this complaint, and if it holds in practice it is the most valuable thing in the 2.0 release.

Note also that pricing for Cloud Computer, Max Ultra, and the Studio tiers was not disclosed in the 2.0 launch materials. Budget for it as an unknown.

Where Manus Falls Short

Blind billing. No pre-run cost estimate remains the single biggest usability flaw.

Long, non-interactive runs. Complex tasks can take hours, and you are largely locked out while they execute. Manus optimizes for hands-off throughput, not conversation.

It needs tight specs. Ambiguous goals produce mediocre output. Manus is poorly suited to creative or open-ended work where you want to iterate toward an answer — it wants a well-defined destination. Studio's edit panel softens this, but does not remove it.

Reliability below frontier-lab standards. Sessions can hang on bot-detection walls and other live-web obstacles. Reviewers have not found it Anthropic- or OpenAI-grade on robustness.

Support quality varies as the product scales.

The Meta Acquisition and the Geopolitical Overhang

This belongs in a buying decision, not just a news roundup.

Meta announced an acquisition of Manus in December 2025, reported at $2–3 billion. Neither party notified Chinese authorities. In January 2026 China's regulators opened an evaluative investigation and issued exit bans on executives. On April 27, 2026, the National Development and Reform Commission blocked the deal, and Meta formally terminated it on June 15, 2026.

Chinese regulators cited three red lines: technology sovereignty, data sovereignty, and national security. The broader precedent — widely described as closing the "Singapore washing" loophole — is that a Chinese-origin startup repapered as a Singapore entity may still fall under Chinese jurisdiction over its data-processing capabilities.

The practical implication: if you work in a regulated industry or handle sensitive corporate data, Manus warrants a genuine vendor review before you hand its sandbox your credentials and connected accounts. For an individual doing market research, this is background noise. For a compliance-bound enterprise, it is not.

How Manus Compares

vs. ChatGPT Agent Mode — ChatGPT Agent is the more accessible all-rounder, bundled into a $20/mo Plus subscription you may already pay for, with Operator-grade browsing and permission checks. Manus goes further with a full virtual computer built for genuinely unattended runs. Choose ChatGPT Agent for convenience and predictable billing; choose Manus when the task should finish without you.

vs. Genspark — The closest direct competitor. Genspark ships deliverables faster and gives you 10,000 monthly credits on its entry plan against Manus's 4,000. Manus goes deeper on hard technical work. Pick Manus when depth beats speed and volume.

vs. Devin — Devin is a specialist software engineer billed in compute units; Manus is a generalist. For pure code, Devin. For research, analysis, documents, and builds under one roof, Manus.

vs. n8n or Zapier — Not really competitors. Those execute workflows you design. Manus decides the workflow itself. Use automation platforms for deterministic, repeated pipelines, and Manus for one-off tasks whose steps you cannot specify in advance.

Who Should Use Manus

Good fit:

• Analysts and consultants producing research reports and market analyses
• Non-technical users who want safe, isolated web automation without scripting
• Founders validating ideas who need a working prototype rather than a spec
• Operations teams with multi-step tasks spanning several connected tools
• Anyone whose bottleneck is hours of tedious execution, not thinking

Poor fit:

• Casual or experimental AI use — credits make browsing expensive
• Creative work requiring tight iteration loops
• Compliance-bound organizations that have not cleared the vendor question
• Teams needing predictable monthly costs
• Developers who want an IDE-integrated pair programmer

The Verdict

Rating: 4.3 / 5

Manus is the most complete expression of the autonomous agent idea currently shipping. The cloud sandbox is not a gimmick — it is what makes genuinely unattended, hours-long work possible, and the real-time replay keeps that autonomy auditable instead of opaque.

Manus 2.0 addresses the right problems. Cascade attacks cost, Studio attacks the un-editable-deliverable problem, and Automations moves the product from scheduled to event-driven. That is a company listening to the correct complaints.

What holds it back is billing you cannot forecast. Launching tasks without a cost estimate, on a plan where one Wide Research run can consume a third of your month, makes Manus stressful to use freely — and freedom to experiment is exactly how you learn what an agent is good at. Until there is a pre-run estimate, budget conservatively.

Start on the free tier. 300 daily refresh credits is enough to learn whether Manus's hands-off model fits how you work. If your tasks are well-specified, multi-step, and currently cost you hours, the $20 Pro plan pays for itself quickly. If your work is exploratory and conversational, ChatGPT Agent Mode will serve you better for the same money.

Pricing and features verified September 29, 2026. Manus 2.0 launched September 28, 2026; pricing for Cloud Computer, Max Ultra, and Studio tiers had not been disclosed at publication. Credit costs and plan details change frequently — confirm on the official pricing page before subscribing.