Manus AI Coding Review 2026: 8 Real-World Tests

TL;DR: Manus AI averaged 4.69 out of 5 across our eight coding tests, with perfect scores on four of them: a Python script, a debugging job, an algorithm explanation and a landing page. It lost points for skipping an explanation, not running code it wrote, and thin accessibility and validation. It is a fast, reliable first draft that still needs a developer to review it.

Manus has changed hands twice since we tested it.

Meta bought the Singapore-based company on 29 December 2025, China's foreign investment security review ordered the deal unwound in April 2026, and Manus resumed independent operations on 1 September 2026.

The coding product grew in the meantime, adding full-stack web apps in October 2025 and mobile apps in December.

Manus AI Coding Scorecard: 8 Tests

Manus scored 5/5 on four of our eight coding tasks and 4 to 4.5 on the rest, for an average of 4.69.

The tests covered Python scripting, debugging, algorithm explanation, front-end pages, a small web app, a code translation, a React scaffold and a browser game.

Task	What it had to do	Score
A. Currency converter (Python)	Convert USD to EUR, INR and GBP with input validation	5.0
B. Debugging find_max (Python)	Find and explain every bug, handle empty and negative lists	5.0
C. Bubble sort explanation	Explain the code line by line for a beginner, with a worked run	5.0
D. Real estate landing page	Hero, six property cards and a contact form, responsive and accessible	5.0
E. Temperature converter web app	Celsius and Fahrenheit conversion with validation and reset	4.5
F. Factorial, Python to JavaScript	Translate with the same input checks and explain the differences	4.5
G. React app scaffold	Four functional components, placeholders and run instructions	4.5
H. Guess Who game	24 characters, yes/no questions, eliminations, final guess, reset	4.0

The 8 Coding Tests, Task by Task

Each task below shows the prompt we used, what Manus returned, and where it lost points.

Task A: Currency Converter in Python, 5/5

The script never crashed. Typing "abc" returned "Invalid input. Please enter a number," and entering -50 or 0 returned "Invalid input. Please enter a positive number," so it checks both the type and the value before converting.

The prompt asked for a program called currency_converter.py that converts a USD amount at fixed rates, validates the input, prints each result to two decimals with its currency code, and shows a sample run for 100:

Write a Python program named currency_converter.py. The program should:
- Take user input: amount in USD (float).
- Convert to EUR, INR, and GBP using fixed rates: 1 USD = 0.92 EUR, 1 USD = 83.5 INR, 1 USD = 0.78 GBP.
- Validate input. If input is not a positive number, print a clear error message.
- Print each converted value formatted to 2 decimals and include currency code.
- Show one sample run with input 100.
Return the full Python code and the sample run output.

With 100 as input it returned 92.00 EUR, 8350.00 INR and 78.00 GBP, each rounded and labeled correctly. The code also carried short comments explaining each part.

Verdict: complete and reliable. It gave the right results, caught all the bad inputs we tried, and explained itself, so it met all six indicators.

Task B: Debugging find_max in Python, 5/5

Manus found all three bugs: a loop that ran one step past the end of the list, a mutable default argument that could leak data between calls, and a starting value of zero that gave the wrong answer for all-negative lists.

We gave it this broken function and asked for a fix, an explanation of each bug, and output for both calls:

def find_max(nums=[]):
    max_val = 0
    for i in range(len(nums)+1):
        if nums[i] > max_val:
            max_val = nums[i]
    return max_val

print(find_max([3,7,2,9]))
print(find_max([]))  # should return None with message "empty list"

The fixed version returned 9 for [3,7,2,9], and for an empty list it printed "empty list" and returned None, as required. We also tried a list of only negative numbers, and it returned the correct maximum.

The fix checks for empty input first, starts from a real value in the list, then compares the rest.

Verdict: thorough. It fixed the function and explained why each bug happened in plain words, which is what a junior developer would need.

Task C: Bubble Sort Explanation, 5/5

Manus traced the example array [64, 34, 25, 12, 22, 11, 90] pass by pass, showing the largest unsorted value moving to the end each time. That trace is what makes the algorithm click for a beginner.

The prompt asked it to explain this code step by step in simple words and add one example run:

def bubble_sort(arr):
    n = len(arr)
    for i in range(n):
        for j in range(0, n-i-1):
            if arr[j] > arr[j+1]:
                arr[j], arr[j+1] = arr[j+1], arr[j]
    return arr

The line-by-line breakdown covered how the function measures the list, loops over it, and swaps neighbors that are out of order. It also explained why the algorithm is called bubble sort.

The task was an explanation, so there was no code to execute and no edge cases to test.

The rest of the document walked through each later pass in the same format, ending with the fully sorted list.

Verdict: a clear, beginner-friendly explanation with a complete trace. It met each indicator that applied.

Task D: Real Estate Landing Page in HTML and CSS, 5/5

Manus added accessibility features we did not ask for: a skip link, visible focus styles and ARIA attributes on the form. It also used semantic header, main and section tags throughout.

The prompt:

Create a single-file HTML page for a real estate landing page. Requirements:
- Hero section with title, subtitle, and CTA.
- Grid of 6 property cards. Each card has a placeholder image, title, price, and short description.
- Contact form with name, email, phone, and message.
- Use accessible HTML (labels, alt text).
- Make layout responsive using CSS only.
- Use placeholder images from https://via.placeholder.com/ (no external assets).
Return the full HTML with embedded CSS.

Opened in a browser, the page rendered cleanly: a hero with title, subtitle and call to action, then six property cards in a grid that reflowed as we resized the window.

The contact form labeled every field, and every image carried descriptive alt text.

Verdict: complete and ready to extend. It met every requirement on layout, responsiveness and accessibility in one self-contained file.

Task E: Temperature Converter Web App, 4.5/5

Manus rejected temperatures below absolute zero with a specific warning, a physical limit we never mentioned in the prompt. Blank or text input returned "Please enter a valid numeric temperature value."

The prompt:

Build a temperature converter web app in one HTML file with embedded CSS and JS. Requirements:
- Input for numeric temperature and dropdown for unit (Celsius or Fahrenheit).
- Convert to the other unit with result shown to 2 decimals.
- Validate input and show friendly error for invalid values.
- Results panel and Reset button.
Return the full HTML file and show sample conversion for 100 C.

The app converted 100°C to 212°F, the results panel updated cleanly, and Reset cleared everything. Labels were tied to their fields, decimals were accepted, and it added Enter-key submission and auto-focus on load.

Verdict: the app works well. It lost half a point because Manus gave no plain-language explanation of the formulas and validation, which the task asked for.

Task F: Factorial Translation, Python to JavaScript, 4.5/5

Manus wrote the JavaScript but never ran it in Node.js or any other runtime, so it could not prove the code worked. It was the only test where the agent skipped execution.

We asked it to convert this function, keep the same behavior and input checks, and explain the differences:

def factorial(n):
    if not isinstance(n, int) or n < 0:
        raise ValueError("Input must be a non-negative integer")
    return 1 if n in [0,1] else n * factorial(n-1)

The JavaScript checked that the input was a number, an integer and not negative, and threw an error otherwise, like the Python version. The recursion and the base cases for 0 and 1 matched, and a sample call with 5 returned 120.

The explanation of type checking, error handling, recursion and syntax was clear.

Verdict: a solid translation, well explained. One verification run would have made it complete.

Task G: React App Scaffold for a Data Science Services Site, 4.5/5

Manus returned the component files and a live link to the running scaffold, so we could check it in a browser before touching the code.

The prompt:

Create a small React app scaffold for a Data Science Services website. Include:
- App.js with a Hero, ServicesList, CaseStudyCard, and ContactForm components.
- Each component as a separate functional component.
- Use simple CSS or Tailwind class names.
- Use placeholder images and sample props.
Return the component code files (App.js, ServicesList.js, CaseStudyCard.js, ContactForm.js) and short run instructions.

The project ran without problems. The hero, services and case study sections loaded with placeholder content, fields were labeled, and images had alt text.

The contact form required name, email and message, but never checked that the email was in a valid format.

Verdict: a strong scaffold with clear setup steps. It needs proper form validation before production.

Task H: Single-Player Guess Who Game, 4/5

The game only understood questions that contained a keyword it knew, such as hair or glasses. Anything phrased more naturally returned a message asking for a supported question type.

The prompt:

Build a single-player Guess Who style game in one HTML file. Requirements:
- Generate a random board of 24 characters with simple attributes (e.g., hair color, glasses, hat).
- Player can ask yes/no questions using a small input or buttons.
- AI answers based on secret character and eliminates choices visually.
- Allow player to make a final guess.
Return the full HTML/JS code and brief notes on how the AI answers questions.

The game ran without errors. Characters displayed, questions worked, eliminations were correct, and Reset restarted cleanly.

Empty input produced a clear prompt, and guesses were limited to clicking a character, which ruled out invalid moves. The notes explained how answers and eliminations worked.

Accessibility was the gap. The HTML used semantic tags but had no alt text, no ARIA labels and no full keyboard navigation.

Verdict: fun and functional, and a success as delivered. Better question parsing and proper accessibility would bring it to full marks.

What Is Manus AI and Who Owns It

Manus AI is an autonomous AI agent that plans and carries out multi-step tasks in its own cloud computer, then reports back when it is done. Butterfly Effect Pte. Ltd., based in Singapore, makes it.

Manus launched on 6 March 2025 as "the first general AI agent" and opened to everyone without a waitlist in May 2025.

It works like one of the autonomous agents a team would otherwise build in-house. It breaks a request into sub-tasks, works through them without step-by-step input, keeps running while you are offline, and sends the result when it finishes.

Ownership has been the story of the past year:

Manus told users on 11 August 2026 that it was separating from Meta to comply with regulators, and that some data created on or after 29 December 2025 would be deleted on 23 and 24 August.

It announced it had resumed independent operations on 1 September. Neither post says who owns the company now.

On models, co-founder Yichao Ji said in March 2025 that Manus used Claude and fine-tuned Qwen models, and the company said Claude 3.7 would run every task.

It has not named its models since, and now sells them only as Manus 1.6 Lite, 1.6 and 1.6 Max.

At launch it reported 86.5% on level 1 of the GAIA agent benchmark, a self-reported figure it has not updated. Our list of Chinese AI agents covers where Manus sits against its rivals.

What Manus Can Build in 2026

Manus now builds and hosts full-stack web apps and mobile apps, which it could not do when we ran these tests.

Manus 1.5, released on 16 October 2025, added web apps with a backend, login, database and custom domains, and Manus tests each app in its own browser.

Manus says 1.5 cut its average task time from 15 minutes to just under four, on its own internal benchmark.

Manus 1.6, released on 15 December 2025, added a Max model and mobile app development. Builds come out as Android app bundles and iOS TestFlight uploads rather than one-click store releases.

• GitHub: export to a private repository with two-way sync, so a developer can review and edit the code outside Manus.
• Plan Mode (July 2026): Manus proposes a plan for approval before it starts building.
• Supabase connector (July 2026): connects an app to an existing Supabase database.
• My Computer (March 2026): works with local files and development environments.
• API: a v2 REST API can create tasks and publish websites from your own code.

These additions target the gaps we found. GitHub sync and Plan Mode give a developer a review point before code ships, which is where Manus lost its points in our tests.

For comparison, see our round-up of AI coding agents for app development and our Cursor vs GitHub Copilot test.

Manus AI Pricing

Manus costs $20, $40 or $200 a month on its pricing page, with 17% off yearly billing. A free tier exists but is limited to the Lite model.

Monthly credits do not roll over, while add-on credits carry forward as long as a paid plan stays active. Hosting a web app built in Manus is billed separately, with a small free monthly allowance.

Pairing Manus with Engineers Who Review Its Code

Manus drafted working code fast in every test, and every point it lost was something a developer would catch in review: missing explanations, untested code, weak validation and accessibility.

Second Talent places vetted full-stack, React and Python engineers from Asia, with AI-tool fluency tested during vetting, at 50-70% of US cost.

Tell us what you are building and we will send a shortlist within 24 hours.

Frequently Asked Questions

Is Manus AI good for coding?

Yes, for first drafts. Manus averaged 4.69 out of 5 in our eight coding tests and produced working code every time, but it needs human review for validation, accessibility and untested code before anything ships.

Who owns Manus AI?

Manus is run by Butterfly Effect Pte. Ltd. in Singapore. Meta acquired it in December 2025, China ordered the deal unwound in April 2026, and Manus resumed independent operations on 1 September 2026. It has not said who owns it now.

Which AI model does Manus use?

Manus has not named its models since March 2025, when it said Claude 3.7 ran every task alongside fine-tuned Qwen models. It now sells them as Manus 1.6 Lite, 1.6 and 1.6 Max.

Is Manus AI free?

There is a free tier with 300 credits a day, up to 1,500 a month, on the Lite model only. Paid plans start at $20 a month for 4,000 credits and access to all three models.