Estimation question

Estimate the compute cost per app built with Replit Agent.

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Estimation of AI compute costs per unit of user output, requiring reasonable assumptions about task complexity and model usage.

State the goal: estimate the average LLM compute cost to build one app end to end with Replit Agent.
Assume an average app takes several iterative agent turns to reach a working state, say 15 turns including planning, code generation, and error fixing.
Assume each turn processes a moderate amount of context, say 5,000 input tokens (code and file context) and 1,000 output tokens.
Apply an illustrative blended price, for example 3 dollars per million input tokens and 15 dollars per million output tokens, clearly marked as an assumption.
Compute: 15 turns times 5,000 tokens equals 75,000 input tokens, and 15 times 1,000 equals 15,000 output tokens per app.
Multiply: (75,000/1,000,000 times $3) plus (15,000/1,000,000 times $15) equals about $0.23 plus $0.23, roughly 45 to 50 cents per app, before compute for code execution and testing.

Breaks the task into discrete agent turns rather than guessing a lump-sum figure.
States all assumptions (turns per app, tokens per turn, price per token) explicitly and separately.
Adds that execution and sandboxing compute (running the generated code) is a separate cost not covered by the token math.
Sanity-checks the final number against what a free-tier user might cost the business, to reason about margin.
Flags that complex apps with many error-correction loops could cost several times this baseline estimate.

Ignoring compute cost for actually running and testing the generated app, not just the LLM tokens.
Picking an unrealistic number of turns without justifying it.

Chapters of the AI PM course that teach what this question tests.