Key Takeaways
- Manus AI announced its return to independence on August 11, 2026, after China’s regulators forced Meta to unwind its acquisition.
- On the GAIA benchmark, Manus scores 86.5% on basic tasks and 57.7% on complex workflows, above OpenAI’s Deep Research tool, but production use reveals server instability, opaque billing and non-deterministic outputs.
- Manus has been through three ownership states in 18 months; the latest independence announcement removes Meta’s oversight without adding anything more transparent on data governance. Three ownership changes in 18 months is not a track record that makes enterprise procurement easy. After China’s regulators forced Meta to unwind its reported $2 billion acquisition, Manus AI announced on August 11, 2026 that it would operate independently again, its third corporate identity since launching in March 2025.
What Manus Actually Does
Manus takes a high-level goal, breaks it into sub-tasks, then executes them autonomously: browsing the web, writing and running code, reading files, querying databases, then delivering a finished output. That puts it in a different category from conversational tools like OpenAI‘s ChatGPT, which respond and suggest rather than act. The “digital employee” framing is apt. You assign the goal, Manus runs the work in a cloud sandbox, and the session continues even after you disconnect.
The engine underneath is a multi-agent setup. Manus orchestrates several models, including versions of Anthropic‘s Claude and Qwen, alongside a browser, code interpreter and file tools, coordinated through a planner-executor-verifier loop. Practical use cases include deep web research with source citations, data analysis and visualisation, content creation and software development. Hand it a CSV with instructions to produce charts and a PDF summary, and it handles the Python, the rendering and the export without further prompting.
On the GAIA benchmark, which tests agents on complex real-world tasks requiring multi-step reasoning and tool use, Manus scores 86.5% on basic tasks, 70.1% on intermediate and 57.7% on complex workflows, above OpenAI’s Deep Research tool on all three levels. Those numbers come from reported benchmark results rather than independent testing, worth noting before treating them as settled.
Where It Breaks Down
The gap between benchmark performance and production reliability is wide. Users regularly hit “server busy” messages, lags and task freezes under load. Running the same prompt twice can produce different outputs, which makes Manus a poor fit for repeatable business automation. The credit billing system compounds this: costs are hard to predict, and billing surprises are a recurring complaint.
Code output quality follows a similar pattern. Prototype-grade generation is genuinely useful; production-grade consistency is not there yet. Generated code ships with bugs, incomplete archives and formatting issues, and debugging still requires a developer who knows what they are looking at. Hallucinations and factual drift in longer tasks mean human review is mandatory for anything compliance-sensitive. For teams already using LangChain or LlamaIndex to build structured agent pipelines, the lack of determinism here is a real friction point.
There is also a specific architectural constraint. Manus relies on API calls and predefined flows rather than native computer vision for high-resolution GUI interactions, so tasks requiring precise visual manipulation of dynamic screen elements accumulate errors and degrade under load.
Enterprise data governance is the sharpest concern. Manus runs in a cloud sandbox operated by a Beijing-based company, with no mandatory compliance checks on what it accesses or acts on. For regulated industries handling client data, financial models or proprietary strategy documents, that architecture requires constant human oversight of agent actions, which cuts into the efficiency gains the platform promises. The August 11 independence announcement removes Meta’s oversight without replacing it with anything more transparent on the governance side.
The Ownership Problem
The corporate history matters for anyone evaluating Manus as infrastructure. Butterfly Effect, a Chinese startup founded in 2022 by Xiao Hong, launched Manus in March 2025 to significant market attention. The company’s earlier product, the Monica browser extension, remains a separate offering. Meta acquired the company for a reported $2 billion in December 2025. China’s National Development and Reform Commission ordered the deal unwound in April 2026, and the platform is now independent again, its third ownership state in roughly 18 months.
That instability is the clearest signal for enterprise buyers. Manus is genuinely capable for ad hoc research, rapid prototyping and exploratory workflows where an imperfect but autonomous agent saves real time. As a foundation for deterministic, auditable business automation, the combination of governance gaps, billing unpredictability and a corporate structure that has shifted twice in 18 months is harder to justify. The account disruptions flagged for the August transition are a concrete illustration of what vendor dependency looks like when ownership changes without warning.
Originally published at https://autonainews.com/manus-ai-returns-to-independence-after-meta-acquisition-unwinds/
