Researched
This analysis is based on documentation, public user reports, and vendor materials — not yet on our own hands-on testing. How we rate
Start with the ownership story, because it's the headline
Most tool reviews don't need to open with regulatory history. This one does.
In December 2025, Meta announced it was acquiring Manus for around $2 billion, describing it as a way to accelerate AI automation across its consumer and enterprise products. Manus had relocated its headquarters from Beijing to Singapore in mid-2025 — the pattern sometimes called Singapore-washing, where Chinese companies establish foreign registration to ease access to capital and reduce geopolitical friction.
Chinese authorities took an interest. In January 2026 the commerce ministry announced a review examining whether the deal complied with export control, technology transfer and overseas investment rules. Reporting during the review indicated the two cofounders had been barred from leaving China.
On April 27, 2026, the National Development and Reform Commission blocked the acquisition and required the parties to unwind it. Reporting suggested the decision was elevated beyond economic regulators to China's National Security Commission, and that officials characterized the deal in strikingly hostile terms.
In August 2026, Manus said it would soon resume operating as an independent company.
The service never stopped running through any of this. But if you're considering building a workflow on Manus, the relevant fact is that this is a company whose ownership was contested by two governments and whose founders were subject to travel restrictions. That's more uncertainty than most vendors carry.
What it actually does
Manus is an autonomous agent rather than a chatbot. You give it an objective and it plans the steps, browses the web in real time, writes and executes code, analyzes files, and delivers a finished artifact.
The difference from asking a general assistant to do something is duration and independence. Manus will work through a task over many steps without checking in, using tools as it goes. A research task that would take you an hour of tab-switching, it handles in the background.
That's the genuine capability. It's also why the pricing works the way it does.
Credits are consumed by work, not by prompts
This is the conceptual shift that determines whether the pricing feels fair or unpredictable.
You don't pay per message. You pay for what the agent does — every page it loads, every script it runs, every file it processes. A simple question costs almost nothing. A deep research task that visits forty sources and writes analysis costs substantially more.
| Plan | Price | Credits |
|---|
| Free | $0 | 300 daily refresh credits |
| Pro | $20/mo | 4,000/month |
| Pro | $40/mo | 8,000/month |
| Pro | $200/mo | 40,000/month |
| Team | from $20/seat/mo | 8,000+/seat, 2-member minimum |
Annual billing saves around 17%.
Two rollover rules worth knowing. Monthly plan credits do not roll over — unused allowance disappears at reset. Purchased add-on credits do carry over, but only while your paid subscription stays active. Let the subscription lapse and you lose them.
The free plan works differently and more generously than you'd expect: 300 credits refreshing every 24 hours rather than a monthly pool, plus one concurrent task and access to the lighter model in agent mode. That's genuinely usable for testing and light work, not a demo.
The Pro ladder is roughly linear
Twenty dollars gets 4,000 credits. Forty gets 8,000. Two hundred gets 40,000.
The first step is exactly linear — double the money, double the credits, no additional features. The $200 tier is where the per-credit rate improves meaningfully, at five times the price for ten times the allocation.
All three Pro tiers carry the same feature set: twenty concurrent tasks, twenty scheduled tasks, access to the full model lineup, and the heavier capabilities like wide-scope research, slide generation and website deployment.
So the choice between $20 and $40 is purely about volume, and there's no reason to jump up until you're regularly exhausting the lower tier. The jump to $200 is for people running production workflows — constant deep research or long-running development tasks — where the improved per-credit economics justify the commitment.
Team starts at $20 per seat with a two-member minimum, and adds what you'd expect: SSO, data training opt-out, usage analytics, internal access controls and shared templates. It's the only plan giving an admin visibility into who's burning credits on what, which matters more than it sounds once several people are running agents.
One caveat: if you subscribed before the rebrand, legacy pricing generally persists until you change plans. Check your invoice rather than the public page.
Where autonomous agents still break
Manus has improved considerably since the wave of hype around its launch, and it's worth being specific about what still doesn't work rather than repeating either the hype or the backlash.
The core problem is compounding error. On a long chain of steps, a wrong turn at step three doesn't announce itself — the agent continues confidently, building on a bad foundation, and delivers a finished-looking result that's wrong in a way that takes real effort to detect. This is not unique to Manus; it's the central unsolved problem in autonomous agents generally.
The second issue is confident completion. The agent will tell you it finished. Verifying whether it actually accomplished what you asked frequently costs more time than the task saved, particularly on anything where correctness matters.
The pattern most experienced users converge on: Manus for research, exploration and first drafts, human review of the output, then a specialist tool for the actual build. Using Manus to research a market and prototype an approach, then Cursor or Lovable to build the thing properly, is a workflow people report working well.
Where it genuinely shines is bounded, multi-source research — the kind of task where the value is in visiting many places and synthesizing, and where an error is visible in the output rather than buried in a chain of dependencies.
How we researched this
We haven't run Manus at volume across a sustained period, and on an autonomous agent the reliability question is exactly what sustained use answers. This page is research.
The pricing and product details come from Manus's published plans and help center documentation, cross-referenced against independent analyses published between April and August 2026 — several of which state they re-checked against the official pricing page, the most recent in July.
The acquisition story is sourced differently and more solidly: reporting from CNBC, NPR, Fortune, The Register and coverage citing Financial Times and Bloomberg, all published around the April 27 decision, with the August independence announcement from CNBC. On a story this consequential we've stuck to mainstream outlets rather than AI-tool blogs, several of which describe the timeline inaccurately.
One source discrepancy worth flagging: Team pricing is reported as both $20 and $40 per seat depending on the analysis and the credit allocation attached. We've gone with the lower figure as the published entry point since that's what the more recently verified sources say, but the tier scales the same way Pro does, so the number depends on the credits you attach to each seat.
What research can't establish: how often Manus gets things quietly wrong on your kind of task. The free tier answers that better than any review can, and it costs nothing.
Predictable budgets. Credit consumption scales with task complexity, which means an unusually ambitious month costs unusually more. If finance needs a fixed number, this model fights you.
Sensitive or client data. Given the company's history and the jurisdictional scrutiny it has attracted, this deserves a deliberate decision. For personal projects it's a non-issue for most people. For anything under an NDA, decide before you start.
Tasks a general assistant handles. ChatGPT and Claude have agentic capabilities, cost $20 flat, and are more polished and more predictable. Manus earns its keep on genuinely long autonomous chains, not on ordinary work.
Anything requiring verified correctness. Compounding error is real. Don't put an unsupervised agent in front of anything where being confidently wrong has consequences.
Production coding. Manus prototypes well and builds poorly compared to purpose-built tools. Use it for the exploration, not the shipping.
Against the alternatives
Against ChatGPT and Claude: both are more reliable, cheaper at a flat rate, and better at the vast majority of tasks. Manus's advantage is specialization in long autonomous execution. That advantage is real and narrow.
Against Genspark: the closest direct competitor on the autonomous agent positioning, with a similar credit model. Worth trialing both on the same task, since the differences show up in execution rather than on feature lists.
Against Devin: Devin is autonomous software engineering specifically, much more expensive and much more focused. If your tasks are coding tasks, that's the comparison to make.
Against CrewAI: CrewAI is a framework for building your own agent systems in Python. Manus is a finished product. If you have engineers and want control, CrewAI. If you want an agent today, Manus.
Against n8n: deterministic workflow automation versus autonomous execution. For repeatable processes where you know the steps, n8n is cheaper and vastly more predictable. For open-ended objectives, that's not what it does.
Against Perplexity: for research specifically, Perplexity is faster, cheaper and better sourced. Manus goes deeper and takes longer. Different tools for different depths of question.
Pricing 2026
| Plan | Price | What you get |
|---|
| Free | $0 | 300 daily refresh credits, 1 concurrent task, lighter model in agent mode |
| Pro | $20/mo | 4,000 monthly credits, 20 concurrent tasks, full model access, wide research |
| Pro | $40/mo | 8,000 monthly credits, same features |
| Pro | $200/mo | 40,000 monthly credits, best per-credit rate |
| Team | from $20/seat/mo | 8,000+ credits/seat, SSO, analytics, access controls, 2-seat minimum |
Checked August 2026. Manus rebranded its former Basic, Plus and Pro tiers into the current Pro ladder in early 2026; legacy subscribers generally retain their original pricing until they change plans. Monthly credits do not roll over; purchased add-on credits carry over only while the subscription stays active. Annual billing saves roughly 17%. Team pricing is reported inconsistently across sources depending on credit allocation. Verify on Manus's pricing page before purchasing.
Use the free tier as a real evaluation. Three hundred credits refreshing daily is enough to learn whether the agent handles your kind of task, at no cost.
Stay at $20 until you're consistently running out. The step to $40 buys volume and nothing else. There's no feature gate to jump for.
Verify output on anything that matters. Confident completion of a wrong task is the failure mode. Budget review time, not just credits.
Decide the data question first. Given the company's regulatory history, this is worth a deliberate answer rather than a default.
Our Verdict
Manus is a genuinely capable autonomous agent and one of the clearer demonstrations of what the category can do when a task is well-suited to it. Long multi-source research, data analysis, prototyping — these are things it does well enough that the credit meter feels reasonable.
The ownership situation is the dominant fact this year. A blocked $2 billion Meta acquisition, a Chinese national security review, founders under travel restrictions, and a return to independence in August. The service kept running throughout, but this is materially more corporate uncertainty than a typical vendor carries, and it belongs in your evaluation.
The pricing is fair and variable. Paying for work done rather than prompts sent is the right model for an agent, and the free tier is unusually generous. The trade is that you can't forecast a month precisely, and the non-rollover rule on monthly credits punishes uneven usage.
Reliability has improved and hasn't been solved. Compounding error on long chains remains the category's central problem and Manus has not escaped it. The realistic use is agent-for-research, human-for-verification, specialist-tool-for-execution.
For individuals with genuinely multi-step tasks and tolerance for a variable meter, the free tier costs nothing and answers the question quickly. For teams, the admin visibility on the Team plan matters more than it initially appears once several people are consuming credits. For anyone handling client or regulated data, the jurisdictional history is a real consideration, not a theoretical one — and that's the question to settle before anything else on this page becomes relevant.
Note: AIVario earns no commission from Manus. This page is based on published documentation, independent pricing research and mainstream news reporting rather than sustained production use.
Best for: Multi-source research and synthesis, data analysis and exploration, rapid prototyping before a proper build, individuals comfortable with usage-based pricing
Not ideal for: Client or regulated data, tasks requiring verified correctness, production software development, teams needing predictable monthly costs, work a general assistant handles fine
Bottom line: A capable agent worth trying free, attached to a company whose ownership was contested by two governments this year. Judge the product on your own tasks; judge the vendor risk separately.
- ChatGPT — more reliable and cheaper for most tasks, with its own agentic features
- Claude — stronger reasoning and writing at a flat monthly rate
- Genspark — the closest direct competitor on autonomous agent positioning
- Devin — autonomous software engineering specifically
- CrewAI — build your own agent systems if you have Python engineers
- Perplexity — faster and better-sourced for research questions specifically
- Lovable — the build tool to pair with Manus prototyping
Frequently Asked Questions about Manus AI
What happened with Meta and Manus?
Meta announced a roughly $2 billion acquisition in December 2025, saying Manus would discontinue operations in China and that there would be no continuing Chinese ownership. Chinese authorities opened a review in January 2026 examining compliance with export control, technology transfer and overseas investment rules. On April 27, 2026, the National Development and Reform Commission blocked the foreign acquisition and required the parties to unwind it. In August 2026, Manus said it would soon resume operating as an independent company. Reporting also indicated that the two cofounders were barred from leaving China during the review.
Is Manus still available and still shipping?
Yes on both counts. The service continued operating throughout the regulatory dispute and has continued shipping features, including updated scheduled task capabilities and new connector integrations. The corporate turmoil has been an ownership story rather than a service interruption. That said, a company that has been through a blocked acquisition, a national security review and founders under travel restrictions carries more uncertainty than most vendors, and that's a reasonable thing to weigh if you're planning to build a workflow around it.
What does Manus cost in 2026?
Free is $0 with 300 daily refresh credits that reset every 24 hours, one concurrent task and access to the lighter model in agent mode. Pro is a ladder rather than a single price: $20 a month for 4,000 monthly credits, $40 for 8,000, and $200 for 40,000. All Pro tiers include twenty concurrent tasks and access to the full model lineup. Team starts at $20 per seat per month with a two-member minimum, adding SSO, access controls, usage analytics and shared templates. Annual billing saves around 17%.
Why did the plan names change?
Manus rebranded its older Basic, Plus and Pro tiers in early 2026, consolidating them under the Pro name with different credit allocations. If you subscribed under the old naming, the help center indicates legacy subscribers keep their original pricing until they upgrade or downgrade — so your renewal may not match the public plan names. This is worth checking against your actual invoice rather than assuming the published lineup applies to you.
How do credits work and do they roll over?
Credits pay for active task processing. Every action the agent takes — loading a page, executing code, analyzing a file — consumes them, which means cost scales with task complexity rather than with the number of prompts you send. Monthly plan credits do not roll over; they reset. Purchased add-on credits do carry over, but only while your paid subscription remains active. Free users get 300 daily refresh credits that reset every 24 hours rather than a monthly pool, which is a different and in some ways more forgiving model.
Is it actually reliable for real work?
Improved, and still not something to run unsupervised on anything that matters. Manus is genuinely capable at multi-step research, data analysis and prototyping — tasks where a human would spend an hour clicking through sources. Where it struggles is the same place all autonomous agents struggle: long chains where an early wrong turn compounds silently, and confident completion of a task it actually got wrong. The practical pattern most users land on is Manus for research and first drafts, then human review, then a specialist tool for the final build.
Should I worry about where my data goes?
It deserves a deliberate decision rather than a shrug. Manus was founded in China, relocated to Singapore in 2025, and has been the subject of a Chinese national security review that concluded its technology remains strategically significant to Beijing. Whatever the corporate registration says, that history is a data governance consideration. For personal projects, learning and public research, most people will reasonably treat it as a non-issue. For client work, proprietary material or anything under an NDA, it's a question to answer before you start rather than after.
Manus or ChatGPT and Claude?
Different shapes of tool. General assistants have agentic capabilities now and are far more polished, cheaper and more predictable for the great majority of tasks. Manus is built specifically around long-running autonomous execution — give it an objective and it works through it over many steps, browsing and running code without asking. When you genuinely have that kind of task, the specialization shows. When you don't, you're paying a credit meter for something a $20 general subscription does more reliably.
