Estimation question
Estimate the compute cost of a single complex Manus task (research plus web build).
- Manus
- Estimation
- Hard
Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.
Start a mock interview on this question · Mock interview from a job description
What this question tests
Estimation of end-to-end compute cost for a complex, multi-modal agent task combining research and generation work.
How to approach it
- State the goal: estimate the LLM and compute cost of one complex task combining web research and building a small output like a simple website.
- Break the task into phases: research (many search and read steps), planning, and generation/build (writing code or content, iterating on errors).
- Assume the research phase involves around 20 tool calls or page reads, each processing roughly 3,000 tokens of context, totaling about 60,000 input tokens plus modest output tokens for summarizing findings.
- Assume the build phase involves around 15 generation and correction turns, each with roughly 4,000 input tokens (accumulated context) and 800 output tokens, totaling 60,000 input and 12,000 output tokens.
- Apply an illustrative blended price, say 3 dollars per million input tokens and 15 dollars per million output tokens, clearly marked as an assumption.
- Sum the two phases: roughly 120,000 input tokens (about 36 cents) plus about 13,000 output tokens (about 20 cents), landing near 60 to 70 cents in LLM cost per task, before adding sandbox/browser execution compute which could meaningfully add to the total.
What a strong answer includes
- Breaks the task into distinct phases (research, build) rather than treating it as one undifferentiated black box.
- States every assumption (tool calls, tokens per call, pricing) clearly and separately, making the estimate auditable.
- Explicitly flags that browser/sandbox execution compute is a real additional cost beyond pure LLM tokens.
- Sanity-checks the total against what a reasonable per-task price to the user might be, to reason about margin.
- Notes that task complexity variance is high, so this is a baseline, not a ceiling.
Common mistakes
- Ignoring non-LLM compute costs like running a browser or code sandbox, which are real for this type of task.
- Giving a single number with no visible breakdown or assumptions.
Likely follow-up questions
- How would this change for a task with far more research steps?
- How much of this cost is sandbox/browser execution versus pure LLM tokens?
- How would you reduce cost without hurting task success rate?
More estimation questions
- YouTube Red is a premium service without ads. Assume that 1.5% of initial YouTube user base signs up for the service. What is the lifetime revenue Google generates from those users?Google · Estimation · Hard
- How would you design a service like Instagram? Estimate server and storage requirement for peak traffic.Google · Estimation · Hard
- How much storage is required to store all of the images on Google Maps?Google · Estimation · Hard
- A company has launched a new drug that eliminates the need for sleep. How would you price this drug?Google · Estimation · Hard
- How much storage space is required to host all the images of Google Street View?Google · Estimation · Hard
- Imagine you created a new type of product that would replace the mobile phone. How would you determine how many units to manufacture?Google · Estimation · Hard
More questions from Manus
Learn the skill behind it
Chapters of the AI PM course that teach what this question tests.
- Chapter 2: Data fluency: SQL, logs, and reading the truth yourself
- Chapter 9: Prove it paid off: outcomes, economics, and pricing
- Chapter 14: Get the job: the AI PM interview loop