Manus is Butterfly Effect's autonomous AI agent, once expected to join Meta AI until China ordered the roughly $2 billion acquisition unwound in April 2026, leaving Manus independent and still running its agents on Anthropic's Claude models rather than its own.
Summary: Manus still turns a single prompt into a finished report or app for $20 to $200 a month, but Meta no longer owns it. Its founders are raising close to $1 billion to buy it back, and the 86.5% GAIA score every review cites has never been independently re-checked.
On April 27, 2026, China's National Development and Reform Commission ordered Meta to unwind its purchase of Manus, the autonomous AI agent it had owned for four months.
Meta had announced the roughly $2 billion deal in December 2025, planning to fold Manus's task-completing agent into its own AI push. The regulator blocked it on national security grounds instead, and by June 11 Meta had cut off Manus's access to its systems and data entirely.
For anyone deciding whether to pay $20 to $200 a month for Manus today, that changes what's actually on offer. It is not a Meta-backed agent with a trillion-dollar company's compliance and support behind it. It is an independent startup, still metering tasks in credits, while its own founders shop for roughly a billion dollars to buy back what regulators ordered Meta to give up.
Manus launched on March 6, 2025, built by Beijing-founded Butterfly Effect, which had already relocated its engineers and headquarters to Singapore. It grew fast: from a beta waitlist of two million users to more than $450 million in annualized revenue by June 2026, according to research firm Sacra. That growth is what made Meta pay up.
The trouble started almost immediately. Chinese officials opened a national-security and export-control review within weeks of the December announcement, and on April 27 the NDRC issued its verdict: unwind the deal, full stop, the maximum penalty the review process allows, according to Forbes. Manus's co-founders were told not to leave China until the review closed. Meta said the deal "complied fully with applicable law." China disagreed, and by June 11 Meta had finished walking away, cutting data sharing and system access rather than contest the order, per Sacra's reporting.
What happens next is still being negotiated, not decided. The founders are trying to raise close to $1 billion from outside investors to buy the company back at or above the $2 billion Meta paid, with their own money filling any gap, Yahoo Finance reported in May. If that financing closes, the plan is a Hong Kong listing through a restructured Chinese joint venture rather than a US one. None of it is signed yet.
Strip away the ownership fight and the product itself hasn't changed. You give Manus a goal, "build a competitor analysis report," say, and separate planning, execution, and research agents split the work, run it in an isolated sandbox, and hand you a finished file while you do something else.
The free tier gives 300 daily credits and one task at a time. Pro Standard runs $20 a month for 4,000 credits, Pro Customizable $40 for 8,000, and Pro Extended $200 for 40,000, with a $40-per-seat Team plan requiring at least two seats.
That's a workable deal for a freelance researcher or a five-person agency that needs one deep report a week and can absorb some unpredictability. It's a worse deal than it looks for anyone budgeting by the task: a complex agent run burns 500 to 900 credits with no cost estimate shown before it starts, and unused monthly credits expire rather than roll over.
Reviewers of Genspark and other autonomous agents describe the same pattern by mid-2026: tasks stalling partway through, branching workflows derailing, and no shared workspace for a team to hand a task to a colleague.
Manus doesn't train its own model. Its agents reason on Anthropic's Claude Sonnet 4 and Opus 4.x line, with a separate fine-tuned Qwen model routing cheaper sub-tasks to keep credit costs down. Manus discloses this in its own documentation, but it's easy to miss while comparing credit prices across vendors.
That has two practical consequences. When Anthropic changes what its models cost or can do, Manus's output quality and Manus's margins move on a schedule Manus doesn't set. And if what you actually want is the freedom to swap models mid-task or shop for the cheapest capable model on a given job, a model-agnostic gateway like OpenRouter gives you that choice. Manus's orchestration layer doesn't.
Nearly every review of Manus leads with the same statistic: 86.5% on GAIA Level 1, ahead of OpenAI's best publicly documented agent score at the time. What most of those reviews don't say is that the number is self-reported and has never been independently re-run since launch in March 2025.
That matters more after April 2026, when a team at UC Berkeley's Center for Responsible Decentralized Intelligence, led by Dawn Song, published a paper showing GAIA's validation answers were sitting in a public dataset on HuggingFace the whole time. Their exploit agent scored roughly 98% on GAIA's 165 questions by reading answers from a local file, without solving a single task, one of eight major agent benchmarks the team broke the same way.
None of that proves Manus's 86.5% was gamed. It does mean the one figure every buyer's guide repeats was measured on a test that leaked its own answer key, and nobody has re-scored Manus since.
The honest rebuttal is that the benchmark was never really the reason to use the product anyway. Closing your laptop and getting a notification hours later with a finished deck or report, something no chatbot competitor does, is directly testable on a free account in an afternoon. You don't need GAIA to tell you whether that works for your workflow. You need one real task and twenty minutes to read the output.
The founders' buyback round is the fact that resolves everything else. If it closes near or above Meta's $2 billion valuation, expect a Hong Kong IPO filing within a few quarters and a company that starts behaving like a normal, if pricey, vendor: SOC 2 attestation is already published, and GDPR certification would plausibly follow a public listing's compliance push. If the round stalls, expect the uncertainty that has kept enterprise buyers away, no confirmed GDPR status, an unsettled cap table, to run into 2027.
Until one of those happens, every dollar spent on Manus is a bet on a company whose ownership is being settled in a Beijing regulator's office and a term sheet, not a boardroom. The free tier is cheap enough to place that bet just to see the output. The $200-a-month plan is not.
No. China's National Development and Reform Commission ordered Meta to unwind its roughly $2 billion acquisition of Manus's parent company on April 27, 2026, and Meta finished cutting operational and data ties by June 11, 2026. Manus's founders are now trying to raise close to $1 billion to buy the company back and eventually list it in Hong Kong. That financing has not closed yet.
Manus's free tier gives 300 daily credits and one task at a time. Pro Standard is $20 a month for 4,000 credits, Pro Customizable is $40 for 8,000, and Pro Extended is $200 for 40,000, with a Team plan at $40 per seat and a two-seat minimum. Complex agent tasks can burn 500 to 900 credits each, and Manus does not show a cost estimate before a task starts.
The score is self-reported by Manus and has not been independently re-run since the company published it at launch in March 2025. In April 2026, a UC Berkeley team showed GAIA's validation answers were publicly available on HuggingFace, letting an exploit agent score about 98% without solving a single task. That does not prove Manus's number was gamed, but it means the benchmark behind the claim has a documented leak nobody has re-tested Manus against.
Manus does not train its own foundation model. Its agents reason on Anthropic's Claude Sonnet 4 and Opus 4.x line, with a separate fine-tuned Qwen model handling cheaper sub-tasks to control credit costs.
Genspark is the closest headline competitor, offering the same kind of autonomous, end-to-end task execution without Manus's ownership uncertainty. For teams that want to pick and swap the underlying model themselves rather than rely on one vendor's choice, a model-agnostic gateway like OpenRouter is worth comparing against either agent.
All AI guides · Browse the AI directory
Still deciding? Get matched.
Smart Match checks every listing in the directory against how you work and what you can spend, then hands you the shortlist and the reason behind each pick.
Start Smart Match