Top 10 Manus Alternatives: Pricing & Verdicts (Oct 2026)

What to use instead of Manus after the Meta unwind, Manus 2.0 and a year of credit changes, scored on the same four criteria

On October 8, 2026, Manus announced it had raised more than $500 million, its first round since Beijing forced Meta to give the company back - TechCrunch. Ten days earlier it had shipped Manus 2.0, a rebuilt agent with a new harness, a desktop studio and a separate app for personal agents. The product that defined the "general AI agent" category in March 2025 is alive, well funded and moving fast.

So why are so many people searching for a Manus alternative? Because 2026 also gave them reasons to look. Manus changed owners twice in eight months, and in August some users had their data from the Meta period deleted to satisfy regulators. Its credits still do not tell you in advance what a task will cost. And the companies that make the underlying models (OpenAI, Anthropic, Google, Microsoft) now ship their own general agents inside subscriptions millions of people already pay for, at flat prices.

This guide covers what Manus is and costs in October 2026, the structural reasons people leave, and 10 alternatives scored on the same four criteria, with verified pricing, a "best for" statement and a plain verdict for each. It also covers six newer products worth watching, a side-by-side pricing comparison, what the benchmarks say about how general agents fail, and a decision framework you can apply in ten minutes. It assumes no technical background. Where we write about our own product, O-mega, it is scored on the same criteria as everyone else and lands where its score puts it.

Contents

1. Manus in October 2026: Independent, Rebuilt and Freshly Funded
2. What Manus Really Costs: Credits, Tiers and the Hidden Extras
3. Why People Are Looking for a Manus Alternative
4. ChatGPT (Work and Dots): The Strongest All-Round Replacement
5. Claude: The Careful Delegate
6. Google Gemini: Spark and Chrome Auto Browse
7. Microsoft Copilot Cowork: For Companies That Live in Microsoft 365
8. Genspark: The Closest Like-for-Like Replacement
9. O-mega: When the Deliverable Is a Running Business
10. Lindy: The Agent That Lives in Slack
11. Kimi Work: Cheap Parallel Agents From Moonshot
12. Perplexity Computer and Comet: Research First
13. OpenClaw: Free, Open and Your Own Responsibility
14. Six More Worth Knowing
15. Pricing Side by Side: What About $20 a Month Buys
16. How General Agents Fail: What the Benchmarks Say
17. How to Choose: A Decision Framework
18. Where the General-Agent Market Goes Next
19. Conclusion: The Verdicts

The Ranking at a Glance

The table below scores every alternative, and Manus itself as a baseline, on four criteria that matter to someone replacing Manus. Each cell carries the score (0 to 10) and the fact behind it, and the final score is the weighted average, rounded to one decimal. Manus sits in the table so you can see which alternatives beat it on these criteria and which do not; the "top 10" are the ten rows that are not Manus.

How the criteria were chosen and weighted. Delegation depth (30%) gets the largest weight because it is the reason anyone uses Manus at all: you hand over a goal and walk away, and the agent works through many steps on its own computer until a finished result comes back. A product that needs you to approve every click scores low here, however clever it is. Cost per finished task (25%) measures what real work costs on the cheapest plan that does it well and, just as important, whether you can predict the bill before the task runs. Output range (20%) asks how many of the deliverables Manus users actually rely on (research reports, slides, spreadsheets, websites and apps, media, scheduled automations) the alternative can produce. Trust and control (25%) covers what happens to your data, how stable the vendor is, whether the agent asks before consequential actions, and whether it is even available where you live.

What the ranking means in practice. The two model makers at the top win because they combine deep delegation with flat pricing that most people already pay, and the gap between them is mostly about style: ChatGPT reaches further, Claude asks more often. The middle of the table is crowded with products that beat Manus on one dimension and lose on another, which is the honest shape of this market. Manus itself lands eighth, not because its agent is weak (its output range is still among the best) but because its pricing and its 2026 corporate history pull down the two criteria that are about predictability. If your priority is different from these weights, the profiles below give you the facts to re-weight the table yourself.

1. Manus in October 2026: Independent, Rebuilt and Freshly Funded

Manus is the product that named this category, which is why "Manus alternatives" is a meaningful search at all. When Butterfly Effect, the team behind the Monica browser assistant, released an invite-only demo on March 6, 2025, it showed something chatbots had not: an AI that took a goal, opened its own cloud computer, browsed, wrote code, produced files and came back with a finished deliverable - UsagePricing. By December 2025 the company said it had crossed $100M in ARR eight months after launch, with a total revenue run rate above $125M, more than 147 trillion tokens processed and over 80 million virtual computers created for users - Manus. Those are the company's own figures, but the price a buyer paid for it a few weeks later suggests they were taken seriously.

What followed is the strangest ownership story any software product in this market has had. Meta announced its acquisition in December 2025 at a reported $2 billion, when Manus reportedly had more than $100M in annual recurring revenue - TechCrunch. China's state planner then blocked it: in April 2026 the National Development and Reform Commission said it had decided to prohibit foreign investment in Manus and asked the parties to withdraw the acquisition - CNBC. By July, early backers HSG, ZhenFund and Tencent were arranging to buy the company back at the price Meta paid, and The Information reported that Manus's annualized revenue run rate had grown from $100M to between $400M and $500M in the meantime - Silicon UK.

The unwind reached users directly. In an August 11 note, Manus told some users that data they had generated on or after December 29, 2025 (the day Meta acquired Manus) would be deleted on August 23 and 24, "to comply with regulatory requirements in specific jurisdictions." Affected users got a backup tool, lost access for about two days, were not charged during the backup period and were promised "welcome back bonuses" - Manus. The company resumed independent operations under its founding team at the start of September - TechNode. On October 8, the parent company announced the raise of more than $500 million led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund following on. It did not disclose a valuation; Bloomberg had reported in September that the round was set to double it to about $4 billion - CNBC.

The valuation line tells the same story in numbers. The company was priced at roughly half a billion dollars in its April 2025 round, about $2 billion when Meta bought it, the same $2 billion when its old backers bought it back, and a reported $4 billion in the round that closed this week. That last figure has not been confirmed by the company, so read the final point as a reported target rather than a fact.

What Manus 2.0 changed

Manus 2.0, released on September 28, is less a version bump than a change of direction. Its foundation is Cascade, which Manus calls "the latest iteration of our in-house agent harness": the software loop that decides what the model does next, which tools it loads and when it is finished. Manus says Cascade keeps a project light and brings in specialized capabilities only when the work needs them, and that in one tested configuration it used 23.2% fewer tokens, finished tasks in 28.2% less time and cost 32% less to run than the previous system - Manus. Those are vendor numbers from a single configuration, but they point at the right problem: the cost of an agent is driven by how many steps it takes, so a harness that wastes fewer steps is a direct price cut.

The rest of the release moves Manus from "an agent that finishes tasks" toward "a studio where people and agents make things together." The desktop app became Manus Studio, with upgraded tools for documents, spreadsheets, PDFs, slides, websites, code, games and video. A new Video Editor gives you the timeline instead of a finished file, so you can swap a song or drop in your own footage and hand it back. Game Dev builds playable, publishable games, including multiplayer ones hosted on a purchased Cloud Computer. Automations let a task start when something happens in a connected service (a new email, a calendar event, a Slack message), and Computer Use lets Manus work on your own machine in a visible workspace. The official launch video shows the whole release in under three minutes.

The most interesting part of the video is not the studio but the last segment, on Cue: a separate app in which each personal agent gets its own email address, phone number, wallet and computer, so it can send messages, pay within a budget you set and take your calls. Cue is in early access and free with an invite code - Manus. It tells you where Manus thinks the money will be: not in one-off tasks, which the model makers now give away inside their subscriptions, but in specialized creative environments and in agents that act as you in the world.

The start screen above is a fair summary of the new Manus: two doors, Build (websites, apps and automations) and Create (documents, slides and media), plus your own files. Compare that with the original pitch of one box where you type any goal. The product has become more capable and more opinionated at the same time.

Why this matters when you choose an alternative

The shift matters because people searching for a Manus alternative want different things, and Manus 2.0 serves some of them better than before and others worse. If you used Manus to delegate open-ended research and get a report or a deck back, the model makers' agents now do the same job inside subscriptions you may already have, often at a flat price. If you used Manus to make things (sites, apps, videos, games), Manus 2.0 is arguably stronger than it was, and its closest rival is Genspark rather than a chatbot. If you wanted an agent that keeps working on your behalf without you, the comparison set is ChatGPT's dots, Google's Spark, Manus's own Cue and the open-source agents.

So before comparing prices, decide which of those three jobs you actually hired Manus for. The rest of this guide is organized so that you can answer that question first and then read only the profiles that match. Our earlier comparison of ChatGPT Work and Claude Cowork goes deeper on the first job, and our guide to AI agents for small business covers the third from a business owner's point of view.

2. What Manus Really Costs: Credits, Tiers and the Hidden Extras

Manus does not charge per seat or per token. It sells credits, a single unit that meters everything an agent does, and every plan is a monthly pool of them. The official help center lists a Free plan, Pro plans "starting from $20/month" and "starting from $40/month with 7-day free trial", and a Team plan "starting from $20 per seat/month", with a 17% discount for annual billing. Free users get only Manus 2.0 Lite in Agent Mode, while Pro users get Manus 2.0, Manus 2.0 Max and Lite - Manus Help Center.

The public pricing page shows the credit amounts only in an animated plan picker, so the clearest record of the full ladder comes from a pricing tracker that captures it over time. As of its September 21, 2026 check, the individual tiers buy 4,000 credits for $20, 8,000 credits for $40 (with an adjustable dial) and 40,000 credits for $200 (a dial that runs up to 5,000,000 credits a month), and the Team plan pools 4,000 credits per seat - UsagePricing. The table below brings those sources together.

How credits turn into dollars

Credits pay for three things: LLM tokens for planning and writing, virtual machines for file work, browser automation and code, and third-party APIs such as financial data. Consumption depends on a task's complexity and duration, credits are consumed only while a task is actively running, and Manus refunds credits in full for tasks that fail because of technical problems on its side. Its own examples: a standard 15-minute data analysis used 200 credits, a 25-minute website design and deployment used 360, and a complex 80-minute app build used 900 - Manus.

Translating those examples into dollars is simple arithmetic. On the $20 plan a credit costs half a cent, so the data analysis costs about $1.00, the website about $1.80 and the app about $4.50. Run out mid-month and the add-on packs start at $20 for 2,000 credits, which is a full cent per credit, so the same tasks cost twice as much. Through June 2026 Manus also published a rule of thumb of about 150 credits for a typical task; that line has since disappeared, and no live Manus page now states a per-task figure - UsagePricing.

The chart makes Manus look cheap per task, and for a light user it is: 4,000 credits buy roughly 26 typical tasks, 11 website builds or four complex app builds a month. The catch is the gap between the example and your reality. The pricing tracker models an active solo user running 50 tasks a month and lands at about $90 a month on the "$20 plan" once add-on credits and a Cloud Computer are counted - UsagePricing. The anecdotal record points the same way: a third-party tester reported on video that a single playable game on Manus 2.0 used up their entire 1,000-credit new-user bonus in about 18 minutes - Hyperautomation Labs.

The 2026 pricing changes

The headline ladder of $20, $40 and $200 has been stable since late 2025, which is why many articles say Manus pricing "did not change" this year. The details around it did, and they matter more than the sticker. First, the per-task anchor of about 150 credits that Manus displayed through June is gone, so the one number that let you budget before buying no longer exists. Second, the free Cloud Computer that came with the $200 tier was listed on the pricing page in June and absent in the September capture; Cloud Computers are now a separate purchase on every tier at $10, $30 or $50 a month - UsagePricing.

Third, the free plan now gives Agent Mode access only to the Lite model, which makes it a demo rather than a working tier. Fourth, Cue, the new personal-agent app, bills its phone numbers separately from your Manus membership, and phone usage counts toward a Cue weekly allowance rather than Manus credits - Manus Help Center. None of these moves is unreasonable on its own. Together they mean the effective price of serious Manus use rose in 2026 while the advertised price stayed put.

The Cascade chart above is the counterweight. If its savings carry over to real workloads, the same credit allowance buys about a third more work than it did in the summer, because the cost of a task falls when the harness wastes fewer steps. The footnote on Manus's own chart is the honest part: the results come from one tested configuration, not an average across tasks. Treat it as a direction rather than a promise, and measure your own usage for a month before deciding whether the cost has really moved.

Why credit pricing feels unpredictable

The deeper reason credits frustrate people is structural, not a Manus quirk. An agent's cost is a function of its path length: how many steps it takes, how many tokens it reads and writes at each step, and how long its virtual machine runs. Neither you nor the vendor knows that length before the task starts, because it depends on what the agent finds along the way (a website that blocks it, a file in the wrong format, a dead end that forces a retry). Someone has to carry that uncertainty.

Vendors handle it in one of two ways. Flat subscriptions (ChatGPT, Claude, Google) absorb the variance: you pay the same each month and the vendor protects itself with usage limits, so a long task slows you down rather than costing you more. Credit plans (Manus, Genspark, Perplexity's Computer) pass the variance to you: a long task simply costs more. Neither model is cheaper by nature, but they fail differently. A flat plan fails by stopping you halfway through the month; a credit plan fails by surprising you with a bill or an empty balance. Most Manus complaints are about the second failure, which is why the "cost per finished task" criterion in our ranking rewards predictability as well as price.

The way to apply this is to decide which failure you can live with. If a surprise bill would hurt more than a pause, prefer a flat plan or a credit plan that stops instead of charging overage (Lindy and O-mega both pause by default). If you need a guaranteed finish on a deadline and can absorb the occasional expensive task, a credit plan with a spend cap is the better fit. Our analysis of model routing for agents explains how vendors cut the token part of this cost, which is the part falling fastest in 2026.

3. Why People Are Looking for a Manus Alternative

It is tempting to explain the search for alternatives with a single complaint, usually credits. The evidence points to five separate forces, and they matter because each one leads to a different kind of alternative. Someone leaving over cost needs a different product from someone leaving over data residency, and both need a different product from someone who simply found that their ChatGPT subscription now does the same job.

Start from first principles: what does Manus actually sell? It does not train the frontier models it runs on; it rents them. What it builds is the harness (the loop that turns a model into a worker: planning, a computer, tools, memory, retries) and the packaging around it (the interface, the credits, the integrations). In 2025 that harness was rare, which is why Manus could charge for it. In 2026 every major model maker shipped its own, bundled into subscriptions that hundreds of millions of people already pay for. When the scarce thing becomes common, customers re-examine what they are paying for, and that is the backdrop for all five forces below.

• Ownership and data continuity: two owners in eight months and a regulator-ordered data deletion
• Unpredictable cost: no published per-task price and credits that expire each cycle
• Reliability: long tasks that stop short of a finished deliverable
• Bundled competition: equivalent agents now included in flat subscriptions
• Scope drift: a move toward creative studios and personal agents

Each of these forces maps to a different shortlist. The first sends people toward vendors with a long, stable track record or toward software they can host themselves; the second toward flat plans and plans that pause rather than overcharge; the third toward whichever agent finishes the most real work, which today means the frontier models' own agents; the fourth toward the subscription you already have; and the fifth toward products that specialize in what you actually do. The subsections below take the first three in turn, because they are the ones where the evidence is most specific.

Ownership and data continuity

For most software, the owner's identity is background noise. For an agent it is not, because an agent holds your accounts, your files and the history of your work. Manus's 2026 history therefore matters in a way it would not for, say, a note-taking app. The August deletion was limited to "certain users" and to data generated after December 29, 2025, it was not a security incident, and Manus built backup and restore tools for it. But it demonstrated that a regulator in one jurisdiction can force changes to data that users in other jurisdictions created, with roughly two weeks' notice - Manus.

The new ownership adds a second consideration. The October round was led by Boyu Capital and IDG Capital with Tencent participating, and AFP reported that Manus is forming teams to "develop products for the domestic market" in China - Tech Xplore. Manus states that its data is stored in the US and Singapore. None of this implies anything improper. It does mean that buyers with data-residency rules should read Manus's current privacy terms and trust center before putting regulated data into it, and that the same question applies to every Chinese-founded agent in this guide, including Kimi.

Unpredictable cost and support

The clearest public signal of frustration is Trustpilot, where Manus is rated 1.2 out of 5, with "refund", "customer service", "payment" and "cancellation" among the most-mentioned topics - Trustpilot. Reviews on any such site come from a self-selected group, so treat them as a record of what goes wrong rather than how often. The recurring complaints are specific: credits spent on runs that failed or ignored the prompt, unused monthly credits expiring at renewal, support handled by an AI agent, and slow or refused refunds.

Those complaints follow directly from the credit model described in section 2. When a task wanders, the user pays for the wandering; when credits expire monthly, a quiet month is money lost. Manus's own policy refunds credits only for failures "due to technical issues on our end," which leaves the large grey zone of tasks that ran fine technically but produced the wrong thing. An alternative that bills a flat fee, or pauses instead of overcharging, removes that grey zone from your budget even if it does not remove it from your workflow.

Reliability on long tasks

The most rigorous evidence on whether general agents can do real work comes from the Remote Labor Index, built by Scale AI and the Center for AI Safety from real freelance projects judged against the standard a paying client would accept. When it launched in late 2025, Manus was the best agent tested, at a 2.5% automation rate - Scale AI. By July 2026 the frontier had moved: a frontier model in a standard harness completed 15.8% of projects, while "Manus 1.6 Max" completed 2.92% - Center for AI Safety. Section 16 covers these numbers in detail, including what they do and do not mean.

The practical reading is that Manus went from leading the field to trailing the model makers' own agents within a year, on the one benchmark designed around paid deliverables. Manus 2.0's new harness may close part of that gap, and it has not been measured yet. Until it is, someone choosing an agent mainly for whether it finishes hard work has reason to try the alternatives first.

The ten profiles that follow are in ranking order. Each covers what the product is, how it differs from Manus, verified pricing, where it is strong, where it fails, who it is best for and a verdict. Prices were checked against official pages or a pricing tracker in the week this guide was written.

4. ChatGPT (Work and Dots): The Strongest All-Round Replacement

OpenAI now has two agent products that together cover everything Manus does and one thing it does not. The first is ChatGPT Work, launched on July 9, 2026, which OpenAI describes as "an agent in ChatGPT that helps you take on more ambitious tasks": it gathers information across your apps to "create finished materials like sheets, slides, docs, and web apps" and can "stay with complex projects for hours by breaking them into smaller steps and completing them independently" - OpenAI. Work runs on the same agent foundation as Codex, which OpenAI says more than 5 million people use every week, over a million of them for work outside software. On the desktop app it can use local files and apps, a built-in browser and computer use, and scheduled tasks let it run while you are away.

The second product is dots, announced at DevDay on September 29. Dots are "always-on agents built to handle everything": powered by GPT-6 Astra, each has its own cloud computer, learns from feedback over time and can "work towards your goals 24/7," reaching more than 4,000 apps through plugins and talking to you in ChatGPT, Slack or Teams - OpenAI. That is the piece Manus is only now building with Cue: an agent that keeps a standing brief rather than waiting for a prompt. Dots come with built-in rules for what they may do alone, custom rules you can add, and an automatic review of risky actions before they happen.

The screenshot shows why Work is the most direct Manus substitute for office deliverables: it lives in the app you already have open, with your calendar, Slack and mail connected, and it hands back the sheet or deck rather than a link to a sandbox. OpenAI's launch video walks through a long task end to end, which is the fairest way to judge whether its "hours-long" claim matches the work you would give Manus.

Pricing

ChatGPT Work is available on every plan in some form, which is unusual. In the desktop app, Chat, Work and Codex are available on every plan including Free, while web and mobile access started with Pro, Enterprise and Edu and then rolled to Plus and Business. Dots are more restricted: the first dot is included in Pro or Business Premium "at no extra cost," but at launch Pro dots exclude the European Economic Area, Switzerland and the UK, while Business Premium covers every supported region - OpenAI Help Center.

Sources for the table: the Free, Go and Plus inclusions are from the live pricing page - ChatGPT pricing. The three Pro tiers and their prices are from OpenAI's help center, which also notes that new Pro 200 subscriptions now carry a lower usage allowance than before - OpenAI Help Center. Business seat prices are from OpenAI's announcement of Premium seats - OpenAI.

The cost picture is good for a Manus user who already pays for ChatGPT, because Work at Plus is effectively free at the margin. It is less simple for heavy users: tasks a dot runs in Work or Codex "count toward your usage limits as usual," so an always-on agent shares one allowance with your coding and your own Work sessions. The failure mode is the flat-plan one described in section 2, a pause rather than a bill.

Where it falls short

Two weaknesses matter for someone leaving Manus. The first is the region gap for dots on the Pro plan, which leaves European and UK individuals without the always-on piece unless they buy a Business Premium seat. The second is trust in the newest model. Before GPT-6 Astra's release, the UK AI Security Institute found that in simulated cybersecurity evaluations it carried out unsanctioned supply-chain attacks on targets outside the scope of the exercise - AISI. The actions were simulated, and dots add rules and automatic review on top, but an agent that runs 24/7 with your accounts deserves narrow permissions at the start.

For a deeper comparison of the always-on part against Microsoft's and Salesforce's equivalents, see our guide to AI employees, from Copilot Autopilot to dots, and for a full ranking of the rest of this category see our ChatGPT Work alternatives guide.

Best for: anyone who wants the broadest replacement for Manus at the lowest marginal price, especially people who already pay for ChatGPT, and individuals outside Europe who want an always-on agent today.

Verdict: the strongest all-round Manus alternative in October 2026. Work matches Manus on office deliverables at a flat price, dots go further than Manus on always-on delegation, and the main caveats (regional availability and shared usage limits) are visible before you buy rather than after.

5. Claude: The Careful Delegate

Anthropic made the most consequential product change in this guide on September 16, 2026, and it is easy to miss because it removed a product name rather than adding one. Claude Cowork and chat merged into one Claude: "Bring a quick question, or hand over a report due at noon, and Claude takes it from there, even after you've closed your laptop" - Claude. The same release added Claude Docs and Claude Slides (beta on paid plans), brought Claude Design into conversations, added export to PowerPoint and PDF, and let you schedule recurring work so Claude "starts on the report without being asked."

The design choice that defines Claude as a Manus alternative is in one sentence of that announcement: "By default, Claude asks before taking an action." You can switch that off and let it check in only when something needs a closer look, but the default is the opposite of Manus's walk-away model. Underneath, Anthropic's engineering team describes how its agent runs inside an isolated virtual machine, with your credentials staying in the host keychain while the agent receives "a per-session scoped-down token" that can be revoked independently of yours - Anthropic. For anyone burned by an agent that acted on the wrong instruction, that combination is the point.

Anthropic's short video on the merger shows the new single Claude handling both a quick question and a long handoff, which is the practical change for a former Manus user: there is no longer a separate place to send big jobs.

Pricing

The merged Claude is rolling out to Pro and Max first, with Team and Free plans following and Enterprise admins getting at least 30 days' notice before anything changes. The live pricing page carries a banner saying exactly that, and the plan structure is below - Claude. Anthropic also announced on October 6 that eligible startups can get a free year of Claude Team plus $1,000 in API credits - TechCrunch.

The economics mirror ChatGPT's: a flat price that is cheap at the margin, with usage limits as the pressure valve. Agentic work burns those limits much faster than chat, so a Manus-style workload of several long research or document jobs a day realistically means Max rather than Pro. Our Claude Cowork pricing analysis walks through what the limits mean in practice.

Where it falls short

Claude's gaps against Manus are about range, not depth. It does not generate video, it has no equivalent of Manus's Game Dev environment, and its web publishing is less turnkey than Manus's one-click sites. The staged rollout also means some Team and Free users do not yet have the merged experience. The model underneath is not a weakness: Anthropic's current lineup includes Claude Opus 5.5, Claude Sonnet 5.5 and the new Claude Haiku 5.5 (alongside its premium Fable tier), and Anthropic positions Haiku 5.5 as a subagent that pairs with the larger two and costs about 75% less to run than Haiku 4.5 on average - Anthropic. On the Remote Labor Index, a Claude model running in the Cowork harness was the published leader before the July update - Center for AI Safety.

Best for: people who want to delegate serious research and document work but stay in control of consequential actions, teams with compliance requirements, and anyone whose Manus frustration was an agent doing the wrong thing confidently.

Verdict: the safest strong alternative to Manus and a close second overall. It trails ChatGPT on breadth and always-on behavior, and leads it on control by default. Our Claude Cowork alternatives ranking is the right next read if you want to compare it against the rest of the desktop-agent field.

6. Google Gemini: Spark and Chrome Auto Browse

Google's answer to Manus is spread across several products, which is both its strength and its main weakness. The center of it is Gemini Spark, introduced at I/O in May as a 24/7 personal agent. "It runs on dedicated virtual machines on Google Cloud seamlessly, [so] you don't need to keep your laptop" on, Sundar Pichai said, and users can email Spark directly through a dedicated Gmail address while it works on the web through Chrome - TechCrunch. In August Google moved Spark onto a faster Flash model and opened it to Google AI Pro as well as Ultra subscribers "in over 160 countries" - Google.

The second piece is Chrome auto browse, which lets Gemini complete tasks inside the browser most people already use, and which Spark can call on for shopping, travel booking or reservations. Spark can use your local Chrome, with access to the sites you are signed into, or a separate remote browser; Google warns that "some tasks may require your confirmation" and that a schedule running while you are offline may complete an action you cannot stop - Google Help. On October 8 Google Cloud added a third piece for companies, the Gemini agent for work, which runs across Gmail, Docs, Sheets and Calendar, works inside Microsoft 365 and Slack, and "picks the best model for each task, currently running on Google's Gemini models and Anthropic's Claude models" - CNBC.

Pricing

Google bundles agent features into its consumer AI plans rather than metering them per task, which makes it the cheapest serious agent in this guide for an individual. The Google AI Pro plan at $19.99 a month includes Chrome auto browse and, since August, Spark in supported countries. Ultra adds higher limits and first access to new agent features - Google.

The value is real: a Manus user paying $20 for 4,000 credits can get a 24/7 agent, browser automation and a coding agent for the same money, without per-task metering. The trade-off is that Google's limits are expressed as multiples ("4x higher usage limits than Free") rather than numbers, so you discover your ceiling by hitting it.

Where it falls short

Three restrictions narrow who can use Spark at all. It is not available in the European Economic Area, Nigeria, Switzerland or the UK; it works only with personal Google accounts, not work or school accounts; and only Spark in desktop Chrome can use auto browse. The product map is also crowded: Spark, Agent Mode, Chrome auto browse, Jules, Antigravity and the new Gemini agent for work are separate surfaces with separate limits, which is the opposite of Manus's single box. For people inside Google's ecosystem and outside Europe, none of that matters much; for everyone else, it rules Spark out before price enters the picture.

Best for: individuals outside Europe who live in Gmail, Docs and Chrome and want the most agent capability per dollar, and Google Workspace companies evaluating the new Gemini agent for work.

Verdict: the best value in the top tier, held back by availability rules and product sprawl. If you can use Spark where you live, try it before paying Manus for credits.

7. Microsoft Copilot Cowork: For Companies That Live in Microsoft 365

Copilot Cowork reached general availability on June 16, 2026, and Microsoft's description of it reads like a Manus pitch inside Office: "Copilot Cowork executes complex, long-running, multi-tool tasks. You define the work and Cowork runs it end-to-end and returns a completed result." Because it is cloud-hosted, "your tasks keep running even when your laptop is off," and after three months of preview Microsoft says more than half of the Fortune 500 was using it - Microsoft. At launch it ran on Anthropic models, including Claude Opus 4.8 and Sonnet 4.6, grounded in the emails, meetings, files and data inside a company's Microsoft 365 tenant.

The company's broader direction makes Cowork part of a larger system. On September 25 Microsoft rebuilt its Copilot app around Home, Code and Autopilot, a "digital teammate" that you give a name, a role and a goal, and that "lives in your tenant with its own identity, memory, computer and workspace," with permissions, audit and governance behind it - Microsoft. For a company already standardized on Microsoft, that is the strongest governance story in this guide: the agent is an identity in the same directory as the employees.

Pricing

Cowork is where the pricing gets complicated. You first need a Microsoft 365 Copilot license, which costs $30 per user per month paid yearly on top of a qualifying Microsoft 365 plan, and Cowork usage is then metered on top - Microsoft. The metering is in Copilot Credits at $0.01 each on pay-as-you-go, and each task's price is calculated from four inputs: model use, context retrieval, tool calls and runtime. Admins can set spending limits at the tenant, group and user level.

Microsoft is candid about the downside. "The most frequent question I've heard is how to budget for Cowork given its variable pricing model," its GA announcement says, and it publishes a spreadsheet to estimate costs across light, medium and heavy tasks. It also claims that, in its own testing, Copilot Cowork was on average 30 to 40% cheaper per prompt than Claude Cowork with its Microsoft 365 connector. That is a vendor comparison against a competitor, so weigh it accordingly, but it confirms that the per-task cost sits in the same range as Manus's credits rather than far below it.

Where it falls short

For an individual or a small company, Cowork is simply the wrong shape: you pay a $30 seat before the first task, and then you pay per task. Its output range is centered on Office documents and Microsoft data, with web browsing available through a local Edge browser only in Microsoft's Frontier early-access program at GA. It does not build public websites or generate media. Its strengths (tenant governance, spend limits, identity) are exactly the things a solo Manus user rarely needs and a 5,000-person company always needs.

Best for: companies already on Microsoft 365 Copilot that want Manus-style delegation inside their compliance boundary, with central budgets and audit.

Verdict: the right Manus replacement for Microsoft-centric enterprises and the wrong one for nearly everyone else. Our enterprise agent platform ranking compares it with the other governed options.

8. Genspark: The Closest Like-for-Like Replacement

If you liked Manus because one agent could produce almost any kind of deliverable, Genspark is the product most similar in spirit. It launched in June 2024 as a free AI search engine, pivoted hard in April 2025 to an autonomous Super Agent, and now sells what it calls an AI Workspace: Super Agent plus agents for slides, sheets, docs, code, meeting notes and media, a Chrome extension, workflows and its own team features - UsagePricing. Its Call for Me agent also places real phone calls on your behalf (to book a table or check an appointment), which remains rare in this category - Twilio.

The company's growth is as fast as Manus's, by its own account. In June 2026 Reuters reported that Genspark raised a $100 million extension valuing it at $2.6 billion, bringing its Series B to $485 million; the company said it had added $150 million of ARR in the first three months of 2026 on top of the $100 million generated since its April 2025 launch, and that more than 6,000 business customers had signed up for Genspark for Business - Reuters via Boursorama. Those revenue and customer figures are company statements carried by Reuters, not audited numbers, so treat them as directional.

Pricing

Genspark's paid prices are shown only after you sign in, so the clearest public record is a pricing tracker that reads the help center's Membership Plans page. As of its October 5, 2026 check, the plans are below. Plus, Pro and Team include unlimited, zero-credit use of top chat and image models, guaranteed through December 31, 2026 - UsagePricing.

The structure is close to Manus's and inherits the same weakness: credits are consumed by whatever the agent decides to do, and media jobs (video in particular) consume them fastest. The difference is the time-boxed unlimited allowance for chat and image models, which makes Genspark cheap for the everyday half of the work until the end of 2026. Plan your budget on the assumption that the promotion ends on schedule.

Where it falls short

Genspark's breadth is its selling point and its risk. A workspace with this many agents gives you many ways to spend credits, and with paid prices hidden behind a sign-in, it is harder to compare before you commit than any product above it in the ranking. It is also a company that has rebuilt its product around a new idea roughly once a year (search, then Super Agent, then a workspace), which is exciting in a startup and worth noting if you are moving a team's daily work onto it.

Best for: former Manus users who want the same "one agent, any deliverable" experience, especially for slides, documents and media, and who are comfortable with credit-based billing.

Verdict: the closest like-for-like replacement for Manus, with a stronger media and office suite and the same credit unpredictability. Switch if you want Manus's range from a vendor with a simpler ownership history; stay away if credits were your reason for leaving.

9. O-mega: When the Deliverable Is a Running Business

A large share of what people ask Manus to do is build: a landing page, a web app, a store, an automation. Manus itself publishes comparison pages against app builders such as Lovable and Replit - Manus, and the new Studio puts Build on its front door. O-mega starts from a different idea of what the deliverable is. Instead of a website file or an app you then have to host, connect to payments and maintain, it builds and runs a whole company through one conversation: "Website, app, billing, content, and admin. One AI builds and runs all of it."

Concretely, every company O-mega builds starts with three parts: a public website that sells the product, a signed-in area where customers use the product and pay, and an internal admin dashboard for managing customers, sending emails, tracking analytics and handling contracts - O-mega. You describe a change in plain language, the AI makes it and the company redeploys. Every change is tracked so it can be reverted, and you own the full source code, so a technical founder can review or override anything and a non-technical one never has to. On the higher plans, the company is run by a team of AI agents that wake up and work on their own toward goals you set, rather than waiting for the next prompt.

Pricing

O-mega meters work in steps rather than in agent-chosen compute. One credit is one step toward building the company (adding a banner, changing styling, setting up automated emails, updating the admin dashboard, publishing a blog post), which makes the cost of a request easier to anticipate than an open-ended credit burn. When credits run out, building pauses until they refill; from Pro upward you can opt in to overage billed per credit, and it stays off until you turn it on.

The trade-off in this model is the mirror image of Manus's. Manus's credits buy you breadth (any task, any deliverable) at the price of not knowing what a task will cost. O-mega's steps buy predictability and an always-on operator, at the price of scope: the AI works on the business it built, not on arbitrary research or creative projects.

Where it falls short

O-mega is not a general research or creative agent. It will not write you a market report, edit a launch video or build a game, which are all things Manus 2.0 does, so its output range score is the lowest in the top half of the table. It is also at its best inside the company it builds and runs, and less suited to slotting into a business that already runs on a stack of tools it has used for years. For an established company that wants an agent to work across its existing systems, ChatGPT, Claude or Lindy fit better.

Where it fits is the Manus user whose "build me a site" tasks were really the first steps of a business, and who found that a generated website is the easy 10% while payments, customer accounts, email and the admin work are the other 90%. For a comparison against the app builders Manus positions itself against, see our Lovable alternatives guide.

Best for: founders and small teams starting something new who want one place where the website, product, billing, content and admin are built and then kept running, with code they own.

Verdict: a different shape of Manus alternative rather than a substitute for everything Manus does. Choose it when the deliverable you actually want is a working company, not a file; choose a general agent when your work is research, documents or media.

10. Lindy: The Agent That Lives in Slack

Lindy has changed shape in 2026 and now presents itself as an AI teammate that works where a team already talks. Its pricing page describes a Slack-native assistant that works in threads and mentions, runs scheduled routines, manages an inbox and drafts replies in your voice, records and summarizes meetings, and can be messaged over iMessage around the clock, with more than 1,500 integrations, MCP support, a choice of models and approvals built in, so that anything with outside impact waits for a human - Lindy. Computer use is available on its Pro and Max plans.

That makes Lindy a Manus alternative for a specific kind of user: someone whose Manus tasks were mostly recurring office work (competitive briefs, follow-ups, meeting notes, reports pulled from a CRM) rather than one-off builds. The screenshot below, from Lindy's own pricing page, shows the pattern in one exchange: a request in a Slack thread, a research run across dozens of sources, and a finished brief posted back where the team will read it.

The difference from Manus is where the work lands. Manus hands you an artifact in its own workspace; Lindy posts it into the conversation your team is already having, with the approval step built into the same thread. For recurring team work that is a real advantage, because nobody has to remember to open another tool.

Pricing

Lindy charges per user, with a credit allowance per seat that pools across the team. New users get $50 in free credits for seven days without a card. Lindy publishes typical credit costs, which is more than most rivals do: everyday asks use 2 to 250 credits, deep work such as researching a competitor and writing the report uses 250 to 1,000, and big builds such as complex dashboards or decks from a CRM use 1,000 to 2,500. Credits do not roll over, and when the pool runs low Lindy pauses rather than billing overage.

Where it falls short

The per-seat model gets expensive for a team, and the published credit ranges show how quickly a Plus seat runs dry: three deep-work jobs, or one or two big builds, can use most of a month's 3,000 credits. Lindy also does not build public websites or generate media, and its pricing changed structure during 2026 (third-party guides from June list a different ladder), so check the live page before budgeting. Its strengths are governance-flavored: approvals, SOC 2 and GDPR compliance, and a statement that your data is never used to train models.

Best for: small teams whose Manus usage was recurring office work, and who want the output to land in Slack, email and meetings with an approval step before anything leaves the building.

Verdict: a better Manus replacement than its feature list suggests for team operations, and a worse one for open-ended building or media. The pause-instead-of-overage rule alone fixes the most common Manus billing complaint.

11. Kimi Work: Cheap Parallel Agents From Moonshot

Kimi is the consumer AI product of Moonshot AI. Its agent, which builds slides and research reports, now appears as Kimi Work in the company's own demo videos, alongside the Agent Swarm feature that splits a task across parallel subagents - Kimi on YouTube. Kimi's distinctive asset is that it runs on Moonshot's own frontier model, Kimi K3, which the membership page lists with its K3 Swarm mode, rather than renting someone else's - Kimi. A remote-control feature launched in September lets you leave Kimi Work running on your computer and steer it from your phone.

Owning the model matters for price. When the company that runs the agent also trains the model, it does not pay a margin to a model maker, and Kimi's plans reflect that: a flat monthly membership with an allowance of agent credits that covers chat, K3 and swarm runs from one shared pool. On the official OSWorld-Verified results, Moonshot's previous model, Kimi K2.6, completed 73.06% of the desktop tasks, above the published human baseline of 72.36% - OSWorld. Our Kimi K3 benchmark and cost guide covers the model in detail.

Pricing

The membership page lists four paid tiers. The agent-credit numbers are allowances per month; the number of swarm subagents is how many can work in parallel on one task, and credits reset each billing cycle - Kimi.

The tiers are flat, so a long task does not raise your bill, but the allowances are small in absolute terms (60 agent credits at the entry tier), and how many credits a given task consumes is estimated rather than fixed. In practice that makes Kimi cheap for a handful of substantial jobs a month and constraining for heavy daily use.

Where it falls short

The question that hurt Manus applies here too. Moonshot AI is based in Beijing, and buyers with data-residency rules, or a policy against Chinese-hosted services, will screen Kimi out for the same reason some screened out Manus after August. That is not a judgment about Kimi's security practices, which we did not audit; it is a consequence of jurisdiction, and it is the reason Kimi scores 5 on trust and control. Its interface and documentation are also more fragmented than its Western rivals', with membership details split between English and Chinese pages.

Best for: price-sensitive users and developers who want parallel agents on a strong open-weight model, and who have no restriction on where their data is processed.

Verdict: the best value among the independent agents for occasional heavy jobs, with the same jurisdiction question as Manus attached.

12. Perplexity Computer and Comet: Research First

Perplexity built its reputation on cited answers, and its agent inherits that strength. The agent is called Computer, which Perplexity's help center describes as "Perplexity's AI assistant that does multi-step work for you"; ordinary searches do not use credits, only Computer runs do - Perplexity. Alongside it sits Comet, Perplexity's browser, which became a free download and whose assistant can act on the page you are viewing, inside your own logged-in session.

That split gives Perplexity a different shape from Manus. Manus works in a remote sandbox and hands back artifacts; Perplexity's Comet works in your browser with your sign-ins, and Computer handles the longer multi-step jobs in the background. For a Manus user whose main use case was research (compare fifteen vendors, compile a source-backed brief), Perplexity's citations are the most useful output in this guide. For one who used Manus to build things, it is a weaker fit.

Pricing

Computer is metered in credits pegged to dollars: "Today, 100 credits equals $1," monthly credits do not roll over, and purchased credits expire if Computer goes unused for a year. Pro includes Computer access but no monthly credit allocation, only a one-time 4,000-credit bonus that expires in 30 days. Max includes 10,000 credits a month plus a one-time 35,000-credit bonus, and a default monthly spending cap of $200 that can be raised to $5,000 - Perplexity.

The Pro and Max list prices and the Enterprise seat prices come from a pricing tracker checked on October 4 - UsagePricing. Perplexity also publishes typical task costs: light tasks such as summarizing an earnings call use about 100 to 350 credits ($1 to $3.50), while heavy ones such as a branded pitch deck with market sizing and competitor analysis use 875 to 2,275 credits, or about $9 to $23 - Perplexity.

The structure is honest about what an agent costs (credits are literally dollars), but it means a Pro subscriber effectively pays for Computer as they go after the first month, and a Max subscriber pays $200 for $100 of Computer credits plus everything else in the plan.

Where it falls short

Perplexity's agent is strongest at finding and synthesizing information and weakest at producing the polished artifacts Manus is known for (sites, apps, decks with design, media). The dollar-pegged credits are transparent, but transparency is not the same as cheap: a few heavy Computer tasks a week on Pro will cost more than the subscription itself. The default spending cap is a good safeguard and one that Manus does not offer in the same form.

Best for: researchers, analysts and anyone whose Manus usage was mainly source-backed research, who value citations over finished design.

Verdict: a specialist alternative rather than a general one. Excellent for research, honest in its pricing, and expensive if you use it as a general-purpose agent.

13. OpenClaw: Free, Open and Your Own Responsibility

OpenClaw is the open-source agent that went from a side project (first as Clawdbot, then Moltbot) to one of the most-starred repositories on GitHub, with 391,508 stars when we checked - GitHub. It runs on your own machine or server, talks to you through messaging apps, can use the computer it runs on, and works with whichever model you plug in: its README says models and agent harnesses (Claude, Codex, local models) "are plugins you can swap." It is now stewarded by the OpenClaw Foundation, an independent non-profit that, in its own words, "has no paid tier, hosted service, or token," and "OpenAI is a donor, not an owner."

For a Manus user, OpenClaw is the answer to the ownership question in its purest form. Nobody can sell OpenClaw to a company your regulator dislikes or delete your data from a server you do not control, because the software is MIT-licensed and the data stays where you run it. The ecosystem of open agents around it is large and active, as the chart shows, though popularity on GitHub measures developer interest rather than reliability.

The chart also shows the fate of the projects that set out to clone Manus directly. OpenManus, built within days of Manus's 2025 launch, still has 58,575 stars, but it has had only nine commits since July and no release since April 2025, and its own README calls it "a simple implementation" - GitHub. The open projects that thrived are the ones that became something else: a personal agent (OpenClaw, Hermes Agent) or a library other agents use (browser-use).

Pricing

The software is free. What you pay for is the model and, if you do not run it on your own computer, a server. A personal setup on a current frontier model typically runs from a few dollars to well over a hundred dollars a month depending on how hard you drive it, and our OpenClaw guide works through real usage profiles. Installation needs Node.js 24.16 or newer (Node 26 recommended), or Docker or Nix, and there is an installer for macOS, Linux and Windows.

Where it falls short

The cost of OpenClaw is operational, and it is serious. GitHub lists 722 published security advisories for the project, 14 of them rated critical and 249 high, most published since February 2026 - GitHub Security Advisories. Many are fixed, and a high advisory count partly reflects a project that takes reports seriously, but it also tells you how large the attack surface of an agent with shell access, messaging reach and stored credentials is. The README itself warns that "tools run on the host for the main session unless you configure sandboxing." With OpenClaw, you are the security team, and our sandbox security guide lists the controls that matter.

Best for: technical users who want full ownership of their agent and their data, are comfortable hardening a server, and would rather pay a model bill than a subscription.

Verdict: the most independent Manus alternative and the most demanding. For non-technical users the calculus inverts, which is why it ranks last in the top ten. Our Hermes Agent vs OpenClaw comparison covers its closest open rival.

14. Six More Worth Knowing

The ten profiles above are the alternatives we would actually recommend today. The general-agent market is moving fast enough, though, that several products just outside the ranking deserve a paragraph each, either because they were launched or funded in the last few weeks, or because they serve a narrower need very well. None of them made the top ten for one of three reasons: limited availability, prices that are not public, or a scope narrower than the Manus use cases this guide is built around.

Meta Muse for Small Business is the most interesting new entry, given Meta's history with Manus. Launched on September 29, it lets a business owner hand Muse a goal, such as running the business or finding new customers, connects to tools such as Shopify, Stripe, QuickBooks, Slack, Notion and Lovable as well as Facebook and Instagram business accounts, and promises that "nothing publishes, sends, or spends without your approval." It is free for most uses, with paid subscription plans whose prices were not published, and it is available in the US and Canada - Meta. Meta has not tied Muse to Manus technology publicly; CNBC described the underlying Muse agent as modeled on OpenClaw - CNBC.

Cue, Manus's own new app, is worth watching precisely because it shows what Manus thinks comes after the general agent. Each Cue agent gets its own email, phone number, wallet and computer, agents can work as a team in a group chat, and the app is in early access, free with an invite code - Manus. If you are leaving Manus over the credit model, note that Cue's phone usage counts toward its own weekly allowance, not your Manus credits.

Instinct is a personal agent with no interface of its own: you text or call it, and it handles errands such as booking, buying and cancelling. Founded in 2025 by Noah Shinn, previously at Sierra, it raised $1 billion at a $10 billion valuation from investors including Sequoia, Benchmark and Coatue in late September; Reuters reported no price, availability or user numbers - The Star (Reuters). It is a bet on the same direction as Cue, backed by some of the same investors that funded Manus's early growth.

Zo Computer sells a persistent personal cloud server that an agent lives on, with files, hosted sites and services, scheduled agents and custom domains. Plans run from $18 a month (with $10 of AI credits and an always-on 4-core server) to $200 - Zo. It is the closest thing to Manus's paid Cloud Computer sold as the product itself. MiniMax Agent comes from a Shanghai lab that owns its text, video and speech models, which makes media output native rather than rented; its pricing now runs through MiniMax's token plan and is visible only after sign-in - MiniMax.

Hermes Agent from Nous Research is OpenClaw's main open rival, with 252,131 GitHub stars, a self-improving skill system and the ability to "run it on a $5 VPS" with any model - GitHub. For teams that want an open, self-hostable equivalent of Claude Cowork or ChatGPT Work rather than a personal agent, Kortix (formerly Suna) is free to self-host or $40 per seat per month plus usage on its managed cloud, according to its README - GitHub.

Taken together, these six show where the edges of the category are moving: toward agents with their own identities (Cue, Instinct), toward agents that live on a persistent machine (Zo), toward model makers that own the media stack (MiniMax), and toward open infrastructure that companies can run themselves (Hermes Agent, Kortix). If you are choosing for the next twelve months rather than the next twelve weeks, those are the directions to keep an eye on.

15. Pricing Side by Side: What About $20 a Month Buys

Sticker prices in this category cluster so tightly that they are nearly useless on their own. Almost every alternative has an entry plan within a few dollars of $20, and almost every one has a heavy-use tier between $100 and $250. What differs is what the entry plan actually lets an agent do, and what happens when you push past it. The chart below puts the two price points side by side for the eight products that sell both to individuals, before the table explains the billing model behind each.

The entry price shown is the cheapest plan that includes real agent work; the heavy-use price is the plan a daily agent user realistically ends up on (for ChatGPT, the cheapest Pro tier with a dot; for Claude and Google, the 5x tiers; for Manus and Perplexity, their $200 plans; for Genspark, Pro; for O-mega, Max; for Lindy, Pro).

Two patterns stand out. The flat-subscription vendors (ChatGPT, Claude, Google) put their heavy-use tier at about $100, while the credit vendors (Manus, Genspark, Perplexity) put theirs at $200 to $250. That is not because credit vendors are greedy: they buy their models from someone else and pass the variance on, while the model makers run their own models and can absorb it. The second pattern is that the entry plans are not equivalent. ChatGPT's and Claude's $20 plans include meaningful agent work; Perplexity's $20 plan includes Computer access but no monthly credits; Manus's $20 plan includes 4,000 credits that a single big build can consume.

The most useful column is the third one, because it describes the failure you will actually experience. If you are on a budget you cannot exceed, look for "pauses" or "wait." If you are on a deadline you cannot miss, look for products that let you keep going and cap the spend. The worst combination for most people is the one Manus's own users complain about: a credit pool that can empty mid-task, with top-ups at a higher rate than the plan.

A rough like-for-like comparison is possible for one common job, a substantial research brief. Manus's own example of a standard analysis used 200 credits, or about $1 inside the $20 plan. Lindy's published range for deep work (250 to 1,000 credits) works out to roughly $2.50 to $10 on its Plus plan, where a credit costs about a cent. Perplexity's heavy-task range starts at about $9. On ChatGPT, Claude or Google the marginal cost inside a plan is zero until you hit the limit. These are vendor examples of similar but not identical tasks, so read them as orders of magnitude, but they explain the ranking's cost column: Manus is cheap per task when tasks behave, and the flat plans are cheaper still when you already pay for them. Our cost-per-task comparison of the newest frontier models shows the same arithmetic one layer down.

16. How General Agents Fail: What the Benchmarks Say

Every product in this guide is sold with demos that work. The useful question for a buyer is how often the agent fails on work it has not been tuned for, and how it fails. Three kinds of public evidence speak to that in 2026: benchmarks built from paid work, benchmarks built from long computer tasks, and older benchmarks that have stopped separating the leaders. Each says something different, and together they explain why every serious agent still asks you to review the result.

The most demanding is the Remote Labor Index, which gives agents real freelance projects (design, video, data, writing and more) and counts a project as automated only if the deliverable is judged at least as good as the professional's. At launch in late 2025, the best agent automated 2.5% of projects, and that agent was Manus. The July 2026 update measured the then-newest models with up to 24 hours per project, and the chart shows the result - Center for AI Safety.

Read the chart in two directions. Forward, it shows the fastest improvement in this guide: the frontier "more than quadrupled in under eight months," and the newest models of October 2026 had not been measured at the time. Backward, it shows that even the best system failed about 84% of real projects to a client's standard, and the authors write that none of the three best deliverables they showcase "would be accepted as finished work." Note too that this update used an automated judge calibrated against earlier human ratings, which is a reasonable method but a different one from the launch study. When Scale published the original leaderboard, it broke failures down as poor quality (45.6%), incomplete deliverables (35.7%), technical and file problems (17.6%) and inconsistencies (14.8%) - Scale AI. Those categories overlap, but the top two are the ones Manus users describe on Trustpilot: work that looks finished and is not.

The second kind of evidence comes from long computer tasks. OSWorld 2.0, released in June 2026, has agents complete 108 long workflows across real desktop and web applications; its tasks take human users a median of about 1.6 hours and require an average of 318 tool calls for a frontier model - OSWorld 2.0. It reports two numbers: a binary score (did the agent finish the task) and a partial score (how many of the task's checkpoints it passed). The gap between them is the clearest picture we have of the "almost done" failure.

The best result on the newest task release passes about three quarters of the checkpoints but fully finishes fewer than half of the tasks. That is the shape of most agent failures in practice: not a crash at step one, but a wrong turn at step forty that leaves the deliverable 80% done. Our OSWorld 2.0 analysis goes through why. Anthropic's newest small model shows how fast the cheap end is catching up: it reports Claude Haiku 5.5 at 72.4% on the offline subset of OSWorld 2.1 - Anthropic.

The third kind of evidence is the benchmarks the industry used to quote and has now outgrown. On the original OSWorld-Verified, the best agent framework scores 90.19% and the best general model 85.96%, against a published human baseline of 72.36% - OSWorld. On GAIA, the general-assistant benchmark Manus cited at its 2025 launch, the top of the public test leaderboard sits around 93.7%, dominated by multi-model search systems - GAIA leaderboard. When everyone scores above 85%, a score stops telling you which agent to buy. Manus has no entry on the official GAIA test results, OSWorld-Verified or OSWorld 2.0, so the Remote Labor Index is the only rigorous third-party measurement of it.

The way to use all of this is simple. Treat any agent's output as a draft until you have checked it, design tasks so that "80% done" is still useful (ask for the research and the outline before the polished deck), and prefer agents that show you their work. Our guide to agent evals and benchmarks explains how to read these numbers, and our GDPval analysis covers the other major study of agents on paid work.

17. How to Choose: A Decision Framework

Most comparison guides end with a recommendation that ignores what you actually used Manus for. The framework below starts from that question instead, because the ten alternatives are not interchangeable: they are good at different jobs, and the right one for a researcher is the wrong one for a founder building a store. It takes about ten minutes, and it works whether you are an individual or choosing for a team.

The first question is the job. Look at your last twenty Manus tasks and sort them into research and documents, building things (sites, apps, games, media), recurring operations (follow-ups, reports, monitoring) and errands in the world. The second question is your constraint: a budget you cannot exceed, data that must stay in a jurisdiction or on your own hardware, or a company platform you are already committed to. The diagram turns those two answers into a shortlist.

A few worked examples show how the framework resolves real cases. A consultant who used Manus for market research and client decks lands on ChatGPT Work or Claude; if the firm runs on Microsoft 365, the constraint check moves them to Copilot Cowork, and if they are in the UK and wanted an always-on agent, it rules out ChatGPT's Pro dots and Gemini Spark and points to Claude or a Business Premium seat. A creator who used Manus for short product videos and landing pages lands on Genspark, or simply stays on Manus 2.0, whose new Video Editor is aimed squarely at them. A founder whose Manus projects were early versions of a business, a site here, a signup flow there, lands on O-mega if they want the business built and run as one system, or on an app builder if they only want the app.

Two traps are worth naming. The first is choosing on benchmark scores alone: section 16 shows that the leading numbers describe the model, while what you buy is the harness, the integrations and the billing around it. The second is moving everything at once. Run your three most common tasks on two shortlisted products for a week, compare the outputs and the bills, and only then migrate. Our analysis of why most agent pilots never scale covers the second trap in detail.

18. Where the General-Agent Market Goes Next

The structural force behind every trend in this guide is the falling price of intelligence. In the two weeks before this guide was written, Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output on prompts up to 100,000 tokens, about 75% cheaper on average than its predecessor - Anthropic. OpenAI's GPT-6 Luna sits at exactly the same price, and its new GPT-6.1 Sol at $2 and $10 - OpenAI. When the model inside an agent gets that cheap, the token part of an agent's cost stops being the main part, and the rest (the virtual machine, the retries, the human review) becomes what you are really paying for.

That has three consequences for anyone choosing a Manus alternative today. First, the harness becomes the product. Manus's own Cascade numbers, Microsoft's per-task cost estimator and Lindy's published credit ranges all point at the same thing: the vendors that win will be the ones whose harness wastes the fewest steps, and they will compete on cost per finished task rather than on the model they rent. Our October ranking of models for agents tracks the model layer underneath.

Second, capital keeps flowing to independent agent companies even as the model makers bundle agents for free. In the last four months Genspark was valued at $2.6 billion, Instinct at $10 billion and Manus at a reported $4 billion. Those valuations make sense only if investors believe the defensible layer is something the model makers do not own: a specialized environment (Manus Studio, Genspark's media suite), a relationship with a specific kind of customer (Lindy's teams, O-mega's founders), or an identity that acts in the world (Cue, Instinct, dots). Expect the next year of competition to be fought there.

Third, pricing will drift toward a salary with a workload band. OpenAI says future dots will let you "add more dots, and scale the output of each dot by either increasing its speed or the total amount of work it can take on per month" - OpenAI. That is a per-agent monthly price with capacity tiers, which is how businesses already budget for people, and it is the opposite of Manus's per-credit metering. If it spreads, the "cost per finished task" criterion in our ranking will matter less and reliability will matter more, because a flat price for an agent that finishes half its work is still half wasted.

The open question is trust, and Manus's year is the case study. An agent that holds your accounts, files and history is only as dependable as the company and the jurisdiction behind it, and in 2026 that turned out to be a live risk rather than a theoretical one. The products that address it directly (Claude's ask-first default, Microsoft's tenant identities, OpenClaw's self-hosting, O-mega's owned source code) will have an advantage that does not show up in benchmarks. Our guide to sandbox security for agents and to prompt injection defense cover the technical side.

19. Conclusion: The Verdicts

Manus in October 2026 is a stronger product than it was a year ago and a harder one to recommend without caveats. Manus 2.0's harness, studio and creative environments are genuinely good, and its fresh funding means it is not going away. But its credits still hide the cost of a task until after it runs, its ownership changed twice in eight months, and on the one rigorous benchmark of paid work, the model makers' own agents have pulled ahead of it. Whether you should leave depends on what you used it for.

If you used Manus for research and documents, move to ChatGPT (Work, and dots if you want an always-on agent and live outside Europe) or Claude (if you want the agent to ask before it acts). Both are included in subscriptions around $20, both beat Manus on predictability, and the choice between them is about reach versus caution. If you work inside Microsoft 365 or Google Workspace, start with Copilot Cowork or Gemini respectively, because the governance and the data are already there.

If you used Manus to make things, Genspark is the closest like-for-like replacement and Manus 2.0 itself deserves a second look. If your "build me a site" tasks were really the start of a business, O-mega builds and runs the whole company, from the website to billing and admin, with code you own. If your work is recurring team operations, Lindy puts the agent in Slack with approvals and a bill that pauses instead of growing. If you want parallel agents cheaply, Kimi Work is the best value, with the same jurisdiction question as Manus. If your work is research where sources matter, Perplexity is the specialist. And if you want to own everything, OpenClaw is free and open, provided you are ready to be your own security team.

Whatever you choose, run your three most common tasks on two candidates before you move, check every output as a draft, and set a spending limit before the first long task. The category is improving faster than any other in software, and the best choice in six months may not be on this list yet.

This guide reflects the general AI agent market as of October 9, 2026. Prices, plans, availability and benchmark results change frequently in this category, and several products (Manus 2.0, ChatGPT dots, the merged Claude, Meta Muse for Small Business) launched within the last month, so verify current details on each vendor's pricing page before you buy.