Manus Review: The Autonomous Agent That's Impressive, Unreliable, and Genuinely Hard to Recommend

What it’s actually good at

Open-web research is where Manus earns its money, and it isn’t close. Give it a real brief (“find the top 12 AI sales tools targeting mid-market SaaS, pull pricing, surface integrations, put it in a spreadsheet”) and the agent opens browsers, runs searches, copies snippets, screenshots its work for verification, and hands you a downloadable file twenty minutes later. For example, “Research the top 10 AI agent platforms launched in 2025, compare their pricing, and create a spreadsheet” is a task Manus can handle with moderate success. That’s a real capability gap with ChatGPT or Claude, both of which can plan the work but can’t actually walk away with it.

The interface is the part nobody talks about enough. You can intervene in multiple ways too. The live Manus’s Computer dashboard streams the AI agent’s actions, so you can see which pages it opens, what searches it runs, which snippets it copies, and even what code or form fills it attempts. You can also switch to a VS Code-style view, where the AI agent’s workspace appears as files, logs, and scripts updating in real time. That observability is what makes Manus feel different from earlier autonomous-agent attempts. You aren’t trusting a black box. You’re watching the work happen, and you can step in the second it heads down the wrong rabbit hole. After a few sessions you stop treating it like a chatbot and start treating it like a remote employee whose screen you can shoulder-surf.

The architecture under the hood is genuinely interesting too. Manus utilizes several AI agents that work together, each specialized in tasks such as browsing the web, analyzing data, or writing code. It can also call different models, such as Claude or Qwen, selecting the best one for each step. The practical upshot: you’re not locked to one provider’s strengths and weaknesses on a given task. The browsing sub-agent uses a vision-capable model, the planning sub-agent uses a reasoning-heavy one, the writing sub-agent uses whatever’s currently best at prose. You don’t manage any of it. It just happens, and the results on heterogeneous tasks are noticeably better than a single-model run.

Scheduled Tasks is the sleeper feature. Scheduled Tasks lets you set Manus to run recurring automations (daily, weekly, or monthly) without needing to re-prompt it each time. I’ve got one running every Monday at 7am that pulls the week’s AI funding announcements, dedupes them against last week’s list, and drops a clean markdown brief in my inbox. That’s a workflow I would never have built by hand, and it’s the kind of thing that justifies the $20 if you can find two or three of them.

Where it falls down

Now the bad news, and there’s a lot of it.

The credit system is the single biggest source of pain in this product, and you’ll feel it within a week. Manus uses a credit-based system where cost per task varies by complexity. Simple questions may consume 10-50 credits, while complex autonomous tasks (deep research, website generation, multi-source data analysis) can burn hundreds or even thousands of credits. On the $20/month Standard plan (4,000 credits), you might complete 20-40 moderate tasks or as few as 4-5 complex ones. The unpredictable credit consumption is one of the platform’s biggest complaints. The dashboard doesn’t tell you what a task will cost before you launch it, so you’re effectively gambling every time you click run. I burned roughly 1,500 credits in a single Saturday morning on three deep-research runs and didn’t notice until I went to start a fourth and the meter was nearly empty.

Reliability is the other rough edge. But in May 2026, the execution is still uneven. Tasks fail mid-stream. Credits burn unpredictably. Complex workflows with branching logic derail. Server reliability during peak hours is poor. I’ve had a research task die at minute 18 of a 25-minute run with no clear error and most of the credits spent. Manus does refund credits when the failure is clearly a platform error, but “clearly” is doing a lot of work in that sentence, and the queue for support is long.

If you came here hoping Manus would replace your code editor, calibrate that expectation down hard. The Manus 1.5 app-building feature generates functional prototypes, not production code. Database schemas are naive, error handling is minimal, security best practices are absent, and architectural decisions are arbitrary. If you need AI-generated code you can actually ship, Devin, Replit Agent, or Claude Code produce significantly better results. Manus’s strength is not coding, it is orchestrating non-coding workflows. That matches my experience exactly. The “build me a landing page” demos are real and the output looks impressive in the preview pane, but the moment you open the generated repo it’s clear no senior engineer would let any of it through code review.

The Meta acquisition is the other thing you need to think about. The footer of the official site now reads ”© 2026 Meta.” Whatever you think of that, it changes the data-privacy conversation. If you’re going to point this agent at client briefs or internal documents, read the Trust Center and run it past whoever owns compliance before you do anything irreversible.

The pricing, read this section before you subscribe

Manus has rebranded its tiers more than once, so check the numbers below against the live pricing page before you pay. As of mid-2026, the lineup looks like this:

The simplified 2026 lineup, listed on the official Manus pricing page, has three plans. Free is $0/month, Pro ladders up through $20, $40, and $200 per month tiers (each with a bigger monthly credit pool), and Team starts at $20 per seat per month with a 2-member minimum.

Free is more usable than most free tiers. The Free plan costs $0 per month and is actually usable, not a watered-down trial. You get 300 daily refresh credits that reset every 24 hours, 1 concurrent task, 2 scheduled tasks, plus access to Chat Mode and Manus 1.6 Lite in Agent Mode. New users also receive a 1,000-credit one-time starter pack, useful for testing one or two real workflows before you decide to upgrade. That’s enough to figure out whether the agent fits your work without putting in a credit card.

The big trap is that the pricing scales linearly. The cleanest way to think about it: $20 of Manus is roughly $0.005 per credit, $40 is also roughly $0.005, and the $200 tier drops the effective rate to $0.005 as well. The pricing is linear, so the only reason to jump tiers is needing more credits than the lower one delivers, not better per-credit value. Don’t upgrade thinking you’re getting a discount. You’re just buying more capacity at the same per-credit rate.

My rule of thumb after six weeks: start free, run three or four real workflows, and only jump to $20 Pro if you’re consistently bumping into the daily limit on something that actually saved you time. Most people will discover that two or three repeatable workflows justify the entry tier and the rest is novelty. Almost nobody who isn’t running production agent pipelines needs the $200 tier.

The bottom line

Manus is the most ambitious product in the AI assistant category right now, and one of the most frustrating. The autonomous research is real, the live dashboard is the best agent UI I’ve used, and the multi-agent architecture is a genuine step beyond plugin-based tool use. But the credit system is a confidence-killer, mid-task failures show up too often, and the code generation isn’t anywhere near where it’d need to be to call this a developer tool.

If you spend serious hours on multi-source research, structured data gathering, or any task that fits the shape of “go away and come back with the answer,” start on Free, find two repeatable workflows that pay back the $20, and upgrade to Pro. If your work breaks neatly into chat-shaped pieces, stay on ChatGPT or Claude. They’re faster, more reliable, and you already know how to drive them. Manus earns a 74: ahead of the field on what it uniquely does, held back from the Editors’ Choice by the rough edges that keep showing up exactly when you don’t have time for them.